Horizontal gene transfer enables programmable gene stability in synthetic microbiota

Abstract

The functions of many microbial communities exhibit remarkable stability despite fluctuations in the compositions of these communities. To date, a mechanistic understanding of this function-composition decoupling is lacking. Statistical mechanisms have been commonly hypothesized to explain such decoupling. Here, we proposed that dynamic mechanisms, mediated by horizontal gene transfer (HGT), also enable the independence of functions from the compositions of microbial communities. We combined theoretical analysis with numerical simulations to illustrate that HGT rates can determine the stability of gene abundance in microbial communities. We further validated these predictions using engineered microbial consortia of different complexities transferring one or more than a dozen clinically isolated plasmids, as well as through the re-analysis of data from the literature. Our results demonstrate a generalizable strategy to program the gene stability of microbial communities.

Introduction

The role that species composition plays in the functions of microbial communities is context-dependent¹. In microbial communities from freshwaters² or soils³, the functional gene repertoires have been shown to be significantly correlated with their species compositions. In many other communities, however, such as those in the human gut⁴, marine waters⁵, macroalgae Ulva australis⁶ and the foliage of wild bromeliads⁷, studies have revealed a marked decoupling between composition and function. In these communities, the fractions of different species vary drastically between habitats, while the functional structure, defined by the relative abundance of functional genes, is remarkably conserved across different communities.

Redundancy, where the same function is shared by distinct species, has been proposed to explain the function-composition decoupling of microbial communities^8,9. The fluctuations of different species sharing the same function compensate for each other. Thus, the relative variations of the shared function will be reduced due to statistical averaging¹⁰ (Fig. 1a and b). This hypothesis was originally derived for eukaryotic ecology; as such, it treated ‘species’ as the basic taxonomic unit assuming every member of the same species is functionally equivalent^6,11. This assumption is not fully applicable to many microbial communities, which are often characterized by a large magnitude of functional and genetic variability among members of the same species and a high level of inter- and intra-species sharing of genetic materials due to horizontal gene transfer (HGT)^6,12. Indeed, the correlation between functional redundancy and functional stability is lacking in some microbial communities. For instance, in soil microbiomes, antibiotic resistance genes are strongly associated with the community’s composition, despite the significant redundancy of the resistance genes³.

Fig.1|. Stability of gene abundance mediated by static or dynamic redundancy.

a. Without redundancy, i.e. the gene being carried merely by a single species, the gene abundance is strongly coupled with the abundance of its host species, exhibiting low functional stability when community composition fluctuates. The ovals of different colors represent different species in the community. The gene is represented by the circular element. The straight line on the right side is displayed for illustrative purpose only and does not represent real data. b. Static redundancy, where the nonmobilizable genes are shared by multiple species, promotes the stability of gene abundance by statistical buffering. The fluctuations of different species hosting the gene compensate for each other, making the gene abundance stable against composition variations. c and d. Depending on HGT rate, dynamic redundancy also promotes the stability of gene abundance. The mobilizable genes are transferred among different species. When the transfer rate is sufficiently high (d), HGT can enable the functional genes to be maintained stably in different communities despite their distinct species compositions. Here, the red arrows represent the HGT route from donors to recipients. The width of the arrows represents the HGT rate.

HGT mediated by mobile genetic elements (MGEs) has been proposed as a contributing mechanism to the function-composition decoupling^3,6,8. That is, the dynamic flow of genes across species might enable the functional genes to be maintained stably in different communities despite their distinct species compositions (Fig. 1c and d). Conceptually this hypothesis is appealing, because HGT is substantial in diverse microbial communities¹². For instance, ~43% of the genes in the pan-genomes of E. coli strains and ~68% in the draft genomes of the human gut microbiome were estimated to be mobilizable^13,14. However, many MGEs are associated with fitness burdens which, in many conditions, can counter their stable maintenance in microbial communities¹⁵. To what extent HGT stabilizes community functions against composition fluctuations remains unknown, due to the lack of quantitative analysis and experimental demonstration.

In this work, we combine theoretical analysis and numerical simulations to show that the stability of gene abundances in microbial communities can be controlled by the rate of gene transfer, which in essence implements dynamic functional redundancy against compositional changes. We experimentally validate such dynamic buffering by HGT in engineered microbial consortia of different complexities. Our results suggest that the decoupling between composition and function is determined by both statistical and dynamical mechanisms and demonstrate a new strategy to control the functional stability of complex communities in a programmable manner.

Results

Theory predicts promotion of gene stability by HGT

To illustrate the basic concept, consider a single gene being carried by a microbial community. The relative abundance of this gene, defined as the fraction of the cells carrying this gene in the community, can depend on the species composition. Let $x_i(i=1,2,\dots ,n$) represent the relative abundance of the gene in a series of parallel communities, each with a different species composition. $n$ represents the total number of communities. The variation of $x_i$ across the communities reflects the stability of gene abundance against composition fluctuations. The smaller the variation, the greater the stability (Fig. 1). If we further assume that the function of the gene is primarily determined by its community-wide abundance, the functional stability is primarily determined by gene abundance stability. That is, our analysis omits the contributions from the gene’s expression levels or activity of gene products. We quantify the degree of gene abundance stability as $\varphi =1/\sigma \left(X_i\right)$, where $X_i$ represents $x_i$ normalized by the mean relative abundance of the gene, and $\sigma \left(X_i\right)$ is the standard deviation of $X_i$ in the parallel communities. High gene stability is characterized by strong decoupling between gene abundances and species composition. Here, we only consider the scenarios where the mean gene abundance across all the communities is not zero. Otherwise, $\varphi$ is undefined.

The stabilization is a purely statistical effect if the gene is non-mobilizable. Dynamic mechanisms will come into play if the gene can be horizontally transferred. To understand how HGT mediates the stability of gene abundance, we considered a simple population of two species transferring a gene by plasmid conjugation. The plasmid carrying the gene can be transferred within the same species or between the two species from plasmid-carrying to plasmid-free cells (Fig. 2a). The plasmid can also be lost due to segregation error¹⁵. Assuming the gene is only carried by this plasmid, the relative abundance of the gene in the community can be represented by the plasmid abundance in the population, i.e., the fraction of plasmid-carrying cells. The community composition is defined by the relative abundances of the two species. To allow the analytical derivation, we assume that the community composition is temporally constant. We also assume that the two species transfer the plasmid at different rates, such that changing community composition will impact the plasmid dynamics. For instance, if species 1 transfers the plasmid more rapidly, a community dominated by species 1 will be more favorable for plasmid persistence than a community dominated by species 2. Varying community composition will result in different plasmid abundances (Fig. 2b), and it can be analytically derived that HGT drives the stabilization of gene abundance in such two-species populations. The stability $\varphi$ increases with HGT rate (Fig. 2c, see Supplementary Information for more details), which can be understood as follows: as the gene transfer rate increases, the response curve of plasmid abundance to species ratio flattens (Fig. 2b), rendering the gene abundance less sensitive to population composition.

Fig. 2|. Theoretical analysis and numerical simulations suggest that horizontal gene transfer (HGT) enables the stabilization of gene abundance against composition fluctuations in two-species communities.

a. A schematic of a community of two species ($s_1$ and $s_2$, green and orange ovals) transferring a single plasmid (black circular elements). The arrows stand for the plasmid transfer from the donors to the recipients. The width of the arrows represents the HGT rate. b. Steady-state plasmid abundance as a function of community composition. It is assumed that species 1 transfers the plasmid twice as efficiently as species 2 and that both species have the same plasmid loss rate. $\epsilon$ is the plasmid transfer rate of species 1 relative to the plasmid loss rate. When the ratio between the transfer rates of species 1 and species 2 is a constant, $\epsilon$ characterizes the overall plasmid transfer rates of the two species. With a relatively small transfer rate ($\epsilon =2$), plasmid abundance varies significantly when the composition transits from $s_1$-dominancy to $s_2$-dominancy, suggesting strong coupling between composition and function. Greater transfer rates ($\epsilon =5$ and $20$) reduce the variability of plasmid abundance and decouple the plasmid abundance from the species composition. c. The gene stability $\varphi$ increases with the relative transfer rate $\epsilon$. We increased the plasmid transfer rate $\epsilon$ from 1 to 6. For each $\epsilon$, we assembled 300 communities with species fractions randomized uniformly between 0 and 1. The steady-state plasmid abundance was calculated for each community. The left panel describes the relationship between the stability $\varphi$ and the relative transfer rate $\epsilon$. The right panels show the species composition and plasmid abundances in the 300 communities, where the plasmid abundances were normalized with their mean values.

HGT promoted gene stability in simple communities

Our theoretical analysis suggests that HGT promotes the stability of gene abundance against composition fluctuations. To experimentally demonstrate such a prediction, we first tested simple synthetic communities of two E. coli strains, MG1655 and Top10, transferring a conjugative plasmid, R388, which is trimethoprim (Tm) resistant (Fig. 3a). The two strains transferred R388 with different rates (Extended Data Fig. 1a). Therefore, the plasmid transfer dynamics might differ in communities with different compositions, allowing the analysis of function-composition decoupling. Top10 expressed Streptomycin (Strp) resistance on chromosome, while MG1655 was Strp-susceptible (Extended Data Fig. 1b and c). We modulated the community compositions using different doses of Strp. Specifically, we constructed 5 communities, each treated by a unique Strp dosage (0, 5, 10, 20 and 40 μg/mL). To control the plasmid transfer rate, we also treated each community with different concentrations of linoleic acid (0, 3 and 8 mM), a known conjugation inhibitor (Extended Data Fig. 1e)¹⁵. The inhibition effect of linoleic acid on R388 transfer was not significantly influenced by Strp doses (Extended Data Fig. 1f). We performed experiments over 15 days undergoing daily dilutions with dilution ratio of 10⁴ or 10⁵ for each community (Fig. 3b). Every 5 days, 200 μl of culture per community was extracted out for measurements of community composition and plasmid abundance. The ratio between MG1655 and Top10 cells was determined by plating on LB and LB+Strp plates (100 μg/mL Strp). The relative abundance of plasmid R388 was determined by plating on the LB and LB+Tm plates (10 μg/mL Tm). Increasing Strp concentrations drastically changed the community composition (Extended Data Fig. 1g). However, the relative abundance of R388 exhibited strong stability against the compositional changes (Fig. 3c). Inhibition of HGT by linoleic acid substantially reduced the stability of plasmid abundances (Fig. 3c). We calculated the values of $\varphi$ in day 5, 10 and 15 under two dilution ratios, and our results suggested that the stability $\varphi$ was promoted by increasing plasmid transfer rate, which is consistent with our theoretical predictions (Fig. 3d). We further carried out a similar experiment, where we mixed Pseudomonas aeruginosa with the two E. coli strains and modulated community compositions by antibiotic treatments (see Methods for more details). The results also confirmed that HGT stabilizes the plasmid abundances against composition variations (Extended Data Fig. 2).

Fig. 3|. Dynamic redundancy by HGT promoted the gene stability of two-strain communities transferring a single plasmid.

a. A schematic of the synthetic communities of two competing E. coli strains, Top10 and MG1655 (green and orange ovals). A single plasmid, R388 (the circular element), was transferred in the communities. Top 10 was resistant to Streptomycin (Strp) treatment, while MG1655 was susceptible. b. The experimental design. 5 parallel communities that transferred the plasmid were assembled, each treated by a unique Strp dosage (0, 5, 10, 20 and 40 μg/mL, respectively). For each community, daily dilutions with dilution ratio of 10⁴ or 10⁵ were performed to maintain the growth. The community compositions and plasmid abundances were measured by selective plating every 5 days from day 0 to day 15. c. The dynamics of R388 abundance during the experiments. The results of three linoleic acid (LAC) concentrations and two dilution ratios were shown here. Data are presented as mean values +/− standard deviations of three replicates. d. The stability of plasmid abundance increased with the plasmid transfer rate. The results of day 5, 10 and 15 were shown in different line styles, while the two dilution ratios were represented by different marker styles.

Our simplified model assumed constant species fractions, while in the experiments the strain compositions are time-dependent (Extended Data Fig. 1g). To evaluate the theoretical prediction in more realistic settings, we updated the two-strain model by accounting for the competition of the two strains and periodic daily dilutions. We parameterized the model by measuring plasmid transfer rates, plasmid burdens, strain growth rates and the drug effects (Strp and linoleic acid) (Extended Data Fig. 1), which allowed us to simulate the temporal dynamics in cocultures of MG1655 and Top10 under different conditions. The simulated R388 abundances and community compositions matched well with our experimental measurements (Extended Data Fig. 3a and b). In particular, the updated model predicted that linoleic acid treatments would reduce the stability, which is consistent with the experimental results and the prediction from the simplified model (Extended Data Fig. 3c). These results suggested that accounting for different complicating factors in the updated model did not change the general relationship between HGT rate and decoupling. Our simplified model captured the basic feature of horizontal gene flow, and the prediction from the simplified model is generally applicable to dynamic communities.

HGT promoted gene stability in complex communities

We have shown that gene transfer promotes the stability of gene abundance in simple synthetic communities. In natural environments, microbial communities are usually complex. They harbor a great diversity of species and MGEs^16,17, the population compositions undergo complex growth dynamics¹⁸, and the MGEs often cause fitness effects on the host cell, either by providing growth benefits or burdens¹⁵. To analyze gene stability in complex communities, we applied a computational framework we previously developed that describes gene flow mediated by multiple plasmids among multiple species (Methods)¹⁹. The framework models the species dynamics by characterizing their growth rates, maximum carrying capacities and the inter-species interactions, and describes selections on plasmids by accounting for their fitness burdens or benefits.

We performed numerical simulations to test the role of HGT in the stabilization of gene abundances in complex communities. We first assembled multiple pools of 100 species transferring 20 plasmids. For each pool, the species-related parameters, including growth rates, maximum carrying capacities and interactions, and the plasmid-related parameters, including plasmid transfer rates, loss rates and fitness costs, were all randomized following uniform distributions in given ranges (Methods). Next, species from the pool were randomly sampled and allocated to 40 local populations (Extended Data Fig. 4a). We assumed that each plasmid was carried by at least one cell of each species. When a species was sampled into a local population, all the plasmids carried by this species were allocated into the population as well. This procedure ensured that every local population transferred all the 20 plasmids. Each local population contained 50 species on average and differed from each other in species compositions. Next, we simulated the population and plasmid dynamics in each local community for 500 hours until the system reached the steady state. We then calculated the plasmid abundances in the local communities to determine the $\varphi$ value of each plasmid. The parameter randomization created a wide range of plasmid abundances, and the abundance stability $\varphi$ was strongly associated with the plasmid transfer rate (Extended Data Fig. 4b and c). When HGT was slow, the plasmid abundances varied drastically across communities, suggesting low stability against composition fluctuations. With high HGT rates, the plasmid abundances were stable, suggesting strong function-composition decoupling (Extended Data Fig. 4b). Here, the steady-state assumption is not critical for the positive correlation between transfer rate and $\varphi$. Numerical simulations indicate the $\varphi$ value was promoted by increasing plasmid transfer rates well before the system reached the steady state (Extended Data Fig. 5).

To experimentally demonstrate the dynamic stabilization by HGT in communities of multiple members, we constructed synthetic microbial communities using E. coli strains from the Keio collection²⁰. The strains are labeled with DNA barcodes, allowing for the quantification of community composition using next generation sequencing (NGS) (Extended Data Fig. 6). We constructed six communities, each consisting of different combinations of barcoded Keio strains. These communities included one community containing 72 strains, two communities each containing 36 strains, and three communities each containing 24 strains (Supplementary Tables 1 and 2). Each community transferred a self-transmissible plasmid isolated from a clinically relevant isolate (denoted as p13) (Fig. 4a)²¹. In each community, p13 was introduced by a single Keio strain (indexed as Keio strain 1). Since different Keio strains transferred p13 with different efficiencies (Extended Data Fig. 7a), community composition was expected to affect the plasmid abundance. We treated each community with different concentrations of linoleic acid (0, 3 and 8 mM), and performed experiments over 15 days undergoing daily dilutions (with dilution ratio of 10⁴ or 10⁵) and regular sampling (every five days) for each community under each condition (Fig. 4b). The plasmid abundances were measured by selective plating, while the community compositions were measured by NGS. The fractions of different strains varied across different communities (Fig. 4c, Extended Data Fig. 7b). The plasmid abundance, however, exhibited different degrees of stability, depending on linoleic acid concentration: higher linoleic acid dose reduced the stability of plasmid abundances (Fig. 4d). This result indicated that HGT enhanced the gene stability against composition variations (Fig. 4e).

Fig. 4|. HGT rate modulated the abundance stability of a single plasmid transferred in synthetic communities.

a. A schematic of the communities of barcoded Keio strains transferring a single plasmid. The green or orange ovals represent different Keio strains. The black circular elements represent the non-transferable plasmid vectors that carry the strain-specific DNA barcodes (shown in green or orange). The red circular element represents the transferable plasmid p13, and the arrow stands for the plasmid transfer from donor to recipient. b. The experimental design. We assembled 6 parallel communities that transferred the plasmid p13, each representing a biologically independent replicate. For each community, daily dilutions were performed to maintain the growth. The community compositions and plasmid abundances were measured by NGS and plating every 5 days from day 0 to day 15. c. The variation of the community compositions at the end of the experiment (day 15). The colored bars indicate different Keio strains, and the heights of the bars represent the relative abundances of the strains within the community. Communities with 0 mM linoleic acid treatment and 10⁴ dilution were provided as an example. d. The dynamics of plasmid abundance during the experiment. The results of three linoleic acid (LAC) treatments and two dilution ratios are shown here. Data are presented as mean values +/− standard deviations of three replicates. e. Reducing HGT rate suppressed the stability of plasmid abundance. The results of day 5, 10 and 15 were shown in different line styles, while the two dilution ratios were represented by different marker styles.

To further examine the prediction in communities of greater complexities, we introduced multiple plasmids into the synthetic communities (Fig. 5a). We constructed six communities, each consisting of different combinations of barcoded Keio strains. Each community transferred 11 self-transmissible plasmids isolated from clinically relevant isolates and 2 other self-transmissible plasmids (RP4 and R388)²¹. These plasmids covered 6 incompatibility groups (IncF, IncI, IncN, IncB/O/K/Z, IncP and IncW), two MOB groups (MOBP and MOBF) and carried distinct types of antibiotic resistance including $\beta$-lactam, tetracycline, aminoglycoside, sulfonamide, macrolide, phenicol and trimethoprim (Fig. 5b, Supplementary Table 3). Plasmid incompatibility is the inability of two plasmids to coexist stably over a number of generations in the same bacterial cell²². Plasmids belonging to the same incompatibility group are incompatible with each other. MOB groups are a plasmid classification system based on plasmid mobility genes²³. The plasmids were introduced by Keio strains 1 or 58 (Supplementary Table 3). We further designed amplicon sequencing primers to allow for the quantification of the plasmids that were introduced into these synthetic communities (Extended Data Fig. 6e, Supplementary Tables 4 and 5) (Methods).

Fig. 5|. Experiments and literature data confirmed the ability of HGT to stabilize plasmid and gene abundances against compositional variations in complex communities transferring multiple plasmids.

a. The design of multi-plasmid transfer experiment. We assembled 6 parallel communities that transferred 13 plasmids, each community with different combinations of Keio strains. Daily dilutions were performed at three different ratios to maintain the growth. The community compositions and plasmid abundances were measured by NGS every 3 days from day 0 to day 15. b. The resistance profiles of the 13 plasmids. Seven types of antibiotic resistances were carried by the plasmids. c. Temporal dynamics of the synthetic communities transferring 13 plasmids throughout the experiment. For each community under a specific dilution ratio (10⁴, 10⁵ or 10⁶), the strain and plasmid abundances were measured by NGS every 3 days, from day 0 to day 15. The abundance of each resistance gene was calculated by summing the abundances of the plasmids that carried this gene. In each panel, the colored bars stand for different Keio strains (left),plasmids (middle), or resistance genes (right), and the heights of the bars represent their relative abundances within the community. d. Plasmid transfer enables the stabilization of plasmid and gene abundances. The abundances of the strains, plasmids and resistance genes at the end of the experiment (day 15) were shown. e. The plasmid abundance stability $\varphi$ increases with plasmid transfer rates. Different colors represent different plasmids, while the marker styles represent different dilution ratios. f. Evaluation of the functional stabilization by HGT using literature data. Data were extracted and reanalyzed from 9 previous studies. In each study, the stability values were calculated based on the plasmid abundances in multiple biological replicates, each with different community compositions. The number of replicates ranged from 3 to 12, depending on the rationale of each study.

The 13 plasmids transferred within the communities differed from each other in their conjugation efficiencies, allowing us to examine the general relationship between gene stability $\varphi$ and HGT rate (Extended Data Fig. 7c). We performed experiments over 15 days undergoing daily dilutions (with dilution ratio of 10⁴, 10⁵ or 10⁶) and regular sampling every 3 days. The community compositions and the plasmid abundances were measured via NGS. The population compositions varied drastically across the parallel communities, due to the different strain combinations (Fig. 5c and d). The plasmids with higher conjugation efficiencies generally exhibited greater abundance stability, confirming the dynamic buffering effects by gene transfer (Fig. 5c–e). The relative abundance of the resistance genes carried by transferable plasmids in the communities also showed high stability against compositional variations, suggesting the capability of HGT to stabilize the community functions (Fig. 5c and d). Our results confirmed the programmable control of the gene abundance stability in microbial communities using gene transfer.

Gene stabilization by HGT is consistent with literature data

The dynamic stabilization by HGT is also reflected by the literature data. In particular, we reanalyzed the data from nine previous studies that provided sufficient measurements on plasmid abundances in communities with different species compositions (Supplementary Tables 6 and 7). These studies covered synthetic communities^15,24,25, soil microbiota^26,27, sludge microbiota²⁸, human intestinal microbiota²⁹ and mouse intestinal microbiota^30,31. In each study, the HGT rate was promoted or suppressed by different factors, such as transfer origins¹⁵, bacteriophages²⁷ or pheromone³¹, allowing the comparison between high transfer rate and low transfer rate. As shown in Fig. 5f, these data confirm that the gene stability increases with HGT rate (see Supplementary Information for more details).

Discussion

Our work demonstrates a simple way to control the stability of gene abundance in complex microbial communities, by broadening the interpretation of functional redundancy. The term ‘redundancy’ has been used to describe the sharing of a gene by multiple species in the same community. It is typically assumed to be static^8,9. HGT makes this redundancy distributed and dynamic, because a member in a community can gain or lose an MGE over time. As long as a sufficient number of members carry a target gene, the stability of the functional genes will be maintained against compositional fluctuations of the community.

Our results provide an explanation for the context-dependent coupling between the species composition and function of microbial communities in different environments. In particular, HGT rates are influenced by various environmental factors, such as spatial structure, nutrient availability or temperature^32,33. The degree of coupling is determined by whether factors and mechanisms exist to strongly promote this dynamic redundancy. For instance, in soil microbiomes where the resistome correlates strongly with community composition, MGEs associated with antibiotic resistance genes are relatively rare, indicating that HGT between soil bacteria is inefficient³. In lakes with high levels of polymetallic pollutants that select for and enrich MGEs, the functions of microbial communities are more decoupled from the species composition, while in lakes without these pollutants, function and composition are significantly coupled^1,34.

Promoting the coupling or decoupling of community composition and functions has broad applications. For instance, in microbiome engineering, multi-species communities integrated with metabolic division of labor are useful to synthesize compounds of medical or industrial importance³⁵. However, the composition of these engineered communities is often susceptible to microbial invasions or fluctuations of environmental conditions³⁶. A strong association between species composition and function can lead to the loss of the desired community functions³⁶. Maintaining the long-term functional stability of these communities presents a major challenge for the application of synthetic microbiota at an industrial scale^36,37. Our work demonstrates that HGT represents an effective strategy to program the functional stability of microbial communities against compositional variations, by generating distributed, dynamic redundancy of target genes. This capability can enable effective optimization of metabolic functions through division of labor³⁸. When doing so, strategies that promote HGT, such as utilizing plasmid vectors with greater transfer rates or broader host ranges, will further stabilize the function of plasmid-encoded genes. In medicine, many pathogenic traits are encoded by MGEs of the microbiota and are insensitive to changes in species compositions³⁹. Such decoupling between composition and community pathogenicity can lead to the persistence of disease states even in the presence of antibiotic treatments that drastically disrupt community composition³⁹. For instance, a virulence plasmid producing enterotoxicity persisted stably in infants with urinary tract infection despite treatment with various antibiotics⁴⁰. A previous study also identified diverse carbapenemase-producing plasmids which persisted through antibiotic treatments⁴¹. Our metric predicts that suppressing the persistence potential of the pathogenic MGEs by reducing their HGT rates might be critical to engineering effective therapeutic strategies.

The uptake and loss of a plasmid can lead to the gain or loss of cellular functions encoded by the plasmid. As such, HGT can cause phenotypic switching in individual cells, which is analogous to that resulting from stochastic gene expression^42,43. Indeed, the mathematical models of these two mechanisms share substantial similarities^42,43. Therefore, studies on plasmids transfer and functional stability can potentially benefit from some prior results from phenotypic switching. For instance, optimal population growth is predicted to occur when phenotypic switching rates match environmental switching rates^43,44. This might imply an optimal plasmid transfer rate in fluctuating environments. In addition, phenotypic switching has been shown to promote the evolutionary adaptation of microbial population to a stressful environment⁴⁵, which suggests a potential role of HGT in community evolvability.

Several caveats need to be considered when applying our metric. Our theoretical analysis and experimental designs did not account for the evolutionary effects caused by de novo mutations. In barcoded communities, mutations on barcodes might bias the measurements of community composition. When distinguishing different strains by antibiotic selection, mutations on resistance markers can also bring a similar bias. These mutations might influence the measurement of composition variations and the calculation of stability. Moreover, the quantification of plasmid abundances can also be affected by mutations that provide antibiotic resistances. Although we estimated the effects of mutations over the 15 days of our experiments to be negligible (see Supplementary Information for more details), measuring function-composition decoupling at a longer time scale (months or years) might need other methods, such as longer barcodes, that are more tolerant to mutation biases. When modulating microbial consortia with a combination of different chemicals, the potential interactions of these chemicals and the consequences of these interactions should also be evaluated when interpreting the experimental results⁴⁶.

Here we assumed that all species expressed the mobilizable genes at the same level. Therefore, the functional stability can be quantified by the stability of gene abundances⁴. However, microbial genomic backgrounds or environmental factors often shape the expression of a specific gene⁴⁷. To examine whether plasmid transfer rate promotes the functional stability when the community members express genes at different levels, we repeated the numerical simulations as shown in Extended Data Fig.4c and randomized the expression level of each copy of the plasmid in each species (see Supplementary Information for more details). Then we quantified the functional stability by calculating the total expression of each plasmid. Our numerical results suggested that increasing plasmid transfer rate consistently promoted the stability regardless of variations of gene expression levels (Extended Data Fig. 8). These results further support the general applicability of HGT in promoting the functional stability of a community.

Our work highlighted the importance of HGT in mediating the abundance stability of mobile genes. However, other mechanisms, especially environmental selection, might also affect function-composition decoupling of microbial communities. The fitness burden of a plasmid and its evolutionary outcomes are often environment-dependent: a burdensome plasmid that eventually gets lost in some environments can be beneficial and maintained in other environments⁴⁸. Environmental selection can also retain the burdensome genes that will be lost otherwise⁴⁹. Conversely, the functions strongly unfavored by selection can eventually get lost regardless of redundancy or HGT⁵⁰. Therefore, the maintenance and stability of plasmids, genes or biological functions in a microbial community is dependent not only on species redundancy or HGT, but also on selective pressures. The quantitative relationship between selection forces and functional stability remains an intriguing question for future studies.

Methods

Strains and growth conditions

The Keio strains were generously provided by Prof. Margarethe Joanna Kuehn at Duke University. Strains containing plasmid barcodes were cultured with 35 mg/ml chloramphenicol to maintain plasmids. Strains containing ESBL plasmids were maintained using 100 mg/ml carbenicillin, or appropriate selection of 10 mg/ml trimethoprim or 100 mg/ml carbenicillin.

Measurement of plasmid conjugation efficiencies

Conjugation efficiencies were estimated using the protocols established by Lopatkin et al. with modifications¹⁵. Overnight cultures of donors and recipients in LB media with appropriate selection agents were resuspended and diluted (1:100) in fresh LB media with antibiotics. Cells were incubated at 37 °C with shaking for 3 hours until they reached exponential phase. The cells were washed twice using fresh LB and resuspended in LB. Donors and recipients were then mixed in 1:1 ratio with a total volume of 200 μL. Mixtures were placed at room temperature (25℃) for 1 hour without shaking. The donor, recipient, and transconjugant densities were measured by diluting the mixtures, and spreading three replicates onto appropriate selective plates. The conjugation efficiency was obtained as $\eta =\frac{T}{D\bullet R\bullet ∆t}$, where T, D, R stand for the cell densities of transconjugant, donor and recipient, respectively.

When measuring the conjugation rates of R388 between Top10 and MG1655, the two strains transformed with R388 served as donors, while the two strains transformed with pJM31 served as recipients. pJM31 is a nonmobilizable plasmid carrying ampicillin (Amp) resistance. Therefore, donors and recipients can be distinguished by Tm or Amp selective plating, respectively, while the transconjugants carry both resistances.

Experiments with MG1655 and Top10 transferring R388

The plasmid transfer experiments were performed in deep well plates (96-well plate, Genesee, Ref#: 27-413S). Single colonies of the two strains carrying and not carrying the plasmid R388 were grown overnight at 37 °C for 16 h with shaking (250 rpm) in LB culture (LB broth from APEX) containing appropriate antibiotics (100 μg/mL Strp, 10 μg/mL Tm). The overnight cultures were resuspended in fresh LB medium, diluted to OD600=0.5 and mixed in 1:1:1:1 ratio. The mixtures were then diluted 1000 folds into a new deep-well plate containing 1 mL of medium with appropriate Strp and LAC dosages. Plates were double sealed with an AeraSeal (Excel Scientific, Ref#: BS-25) and a Breathe-Easy Film (USA Scientific, Ref#: 9123-6100) membrane and grown in a microtiter plates shaker at 700 rpm at 37°C. Every 24 hours, each community was diluted (${10}^4$ and ${10}^5$ dilution ratio). Communities were sampled every 5 days. The ratio between MG1655 and Top10 cells was determined by plating on LB and LB+Strp plates (100 μg/mL Strp). The relative abundance of plasmid R388 was determined by plating on the LB plates and LB+Tm plates (10 μg/mL Tm).

Experiments with Pseudomonas aeruginosa, MG1655 and Top10 transferring R388

Experiments were performed in deep well plates (96-well plate, Genesee, Ref#: 27-413S). Single colonies of P. aeruginosa (strain PA14, carrying a Cm resistant plasmid) and MG1665 (carrying Tet resistance) were grown overnight at 37 °C for 16 h with shaking (250 rpm) in LB culture (LB broth from APEX). Top10 (Strp resistant) cells containing the R388 plasmid were culture overnight with appropriate antibiotics to maintain R388 (10 μg/mL Tm). The overnight cultures were resuspended in fresh LB medium and growth to a minimum OD600=0.3 before being mixed at a 1:1:1 ratio. The mixtures were then diluted 500 folds into a new deep-well plate containing 0.5 mL of medium with different antibiotic (none, Cm, Tet, Strp or Strp+Tet) to modulate population composition and LAC dosages to modulate plasmid transfer rates. Plates were double sealed with an AeraSeal (Excel Scientific, Ref#: BS-25) and a Breathe-Easy Film (USA Scientific, Ref#: 9123-6100) membrane and grown in a microtiter plates shaker at 700 rpm at 37°C. Every 24 hours, each community was diluted (${10}^4$ and ${10}^5$ dilution ratio). Communities were sampled every 5 days and underwent dilutions and selective plating in order to quantify population composition and plasmid abundance.

The relative abundance of MG1655 and Top10 cells was determined by plating on LB+Tet (10 μg/mL Tet) and LB+Strp plates (100 μg/mL Strp), respectively. Total CFU was determined by plating on LB, and PA14 was quantified by removing the abundance of MG1665 and Top10 from the total population abundance. The CmR plasmid carried by PA14 was not used for quantification by plating because it might be lost in communities without Cm selection. Cm resistance was only used to promotes PA14 abundances in one of the five communities. The relative abundance of plasmid R388 was determined by plating on Tm selective plates.

Isolation and characterization of ESBL plasmids from pathogens

A selection of ESBL plasmids were isolated from E. coli pathogens collected from patient bloodstream infections at Duke University Hospital over 2002 to 2014. All plasmids selected for this study were sequenced in previous work²¹ and plasmid sequences were used to design plasmid-specific primers for PCR genotyping. Plasmids were transferred to Keio strains by conjugation. Briefly, conjugation was performed by growing overnight cultures of the donor (ESBL pathogen from the Duke University Hospital) and recipient (a selected strain from the Keio collection already containing a plasmid barcode) strains for 16 hours (LB broth, 37°C, 225 rpm) with appropriate selection (carbenicillin for donors and chloramphenicol for recipients). Overnight cultures were mixed together at a 1:1 ratio and diluted two-fold in fresh LB medium. Cells were grown in co-culture for 1 hr at 37°C. Cells are then diluted and plated with dual-selection plating (chloramphenicol and carbenicillin) to select for transconjugant cells (barcoded Keio strains containing an ESBL plasmid). Colonies were plated and grown overnight at 37°C. Single colonies from each experiment were then picked and PCR-genotyped. PCR products were visualized by gel electrophoresis and samples showing a positive band were validated by Sanger sequencing for both strains (using the strain-specific barcode) and plasmid identity.

The plasmids genotypic resistance profile was determined from sequencing data, as previously reported (Table S2 of the refered publication⁵¹). Briefly, a database of 247,822 protein sequences (from the National Center for Biotechnology Information⁵²) and 2423 antibiotic resistance proteins (from ResFinder database, retrieved 22 May 2018 (http://www.genomicepidemiology.org)⁵³ as predicted by prodigal v2.6.3⁵⁴) was compiled and used to annotate genomes using Prokka v1.13⁵⁵.

Construction of barcoded Keio strains

Barcoded plasmids was generated following the previously published protocols⁵⁶. Briefly, a plasmid vector was linearized (p15A-mCherry-Chloremphenicol, Extended Data Fig. 5a) by PCR amplification (Supplementary Table 4, Primers 1 and 2, Q5^® High-Fidelity 2X Master Mix, New England Biolabs, Ref#: M0492)). Gibson assembly cloning was used to introduce a synthetic barcode sequence (synthesized by an external source) into the plasmid backbone. By design, synthetic barcodes contain two random sequences of 18 base pairs surrounded by Illumina sequencing adapters that serve as targets for PCR amplification and generation of an amplicon library for sequencing. The products of Gibson Assembly reaction were transformed into electrically competent cells⁵⁷. Transformed cells were grown for 16 hours with selection and the plasmid library was harvested using a plasmid miniprep kit (Qiagen Plasmid Mini Kit, Ref#: 12143). Individual strains were barcoded by chemical transformation using the plasmid library⁵⁸. The barcode identity of each strain was confirmed by Sanger sequencing (Supplementary Table 4, Primers 3 and 4).

Experiments with barcoded Keio strains and transferable plasmids

Experiments were performed in deep well plates (96-well plate, Genesee, Ref#: 27-413S). Each strain was streaked on LB plates with selection and single colonies were used to inoculate overnight cultures in LB medium that were grown for 16 hrs. At 16 hours, strains were diluted 50-fold in 1 mL of medium without selection and allowed to grow for 2-3 hours until all strains reached a minimum OD600 of 0.3. The additional dilution step before constructing complex communities was performed to minimize the variability in initial densities of different strains. Complex communities were constructed by mixing selected strains at equal volumes in a mixing trough, homogenizing the mixed culture by gently pipetting, and diluting cells 200-fold into a new deep-well plate containing 1 mL of medium with appropriate selection or treatment conditions. Strain 1 introduced to complex communities contained all the clinical or mobilizable plasmids. Plates were double sealed with an AeraSeal (Excel Scientific, Ref#: BS-25) and a Breathe-Easy Film (USA Scientific, Ref#: 9123-6100) membrane and grown in a microtiter plates shaker at 700 rpm at 37°C. Every 24 hours, each community was diluted (${10}^4$ to ${10}^6$ dilution ratio). Communities were sampled (every 3 days) by extracting 200 μl of well mixed culture. Samples are centrifuged for 5 min at 16,000 rpm to form a cell pellet. Cell supernatant was removed, and cell pellets were stored at −20°C until downstream processing for NGS.

NGS library preparation

The NGS library was prepared following the previously published protocols⁵⁶. The DNA concentration of each sample was normalized based on OD600 measurement by resuspending cell pellets stored at −20 °C in nuclease-free water (Corning, Ref#: 46-000-CM) such that a 2 μL sample of cell lysate should contain approximately 1 million cells (assuming 1 OD600 = $8\times {10}^8$ cells/mL). Cell pellets were resuspended in nuclease-free water, resuspended by pipetting and lysed by boiling at 98 °C for 20 minutes. Lysed cells were pelleted by centrifuged (16,000 RPM for 5 min) and cell lysate was removed and stored at −20 °C.

NGS amplicon library preparation was performed using PCR to ligate appropriate sequencing primers, adapters, indexes and unique molecular identifiers. Strain-specific barcodes introduced on non-transmissible plasmids were designed to incorporate NGS sequencing adaptors. For ESBL plasmids, an additional PCR reaction was required to isolate the plasmid-specific DNA sequence to act as a barcode. Briefly, a set of primers were designed for each plasmid to amplify a unique portion of DNA, generating a 93 base-pair plasmid-specific ‘barcode’. In order for primers to be compatible with a pooled-PCR protocol, primers were design to have compatible annealing temperatures (63-65°C) and selectively amplify only in the target plasmid. Negative control experiments confirmed that each primer set only generated a PCR amplicon in the target plasmid. Positive control experiments confirmed that the DNA band generated by the pooled PCR protocol was representative of the standard PCR protocol containing only one primer set amplified at the optimal annealing temperature. In the pooled PCR protocol, each primer was diluted to a final working concentration of 500 nM. Primer sequences for each plasmid are provided in Supplementary Table 5.

A two-sided size selection was performed using magnetic bead to sample ratios of 1X and 0.8X. By design, plasmid-specific primers used in the pooled PCR protocol introduced Illumina sequencing adaptors such that the plasmid ‘barcodes’ mimic the structure of strain barcodes after the pooled PCR. After magnetic bead clean-up, the cleaned plasmid barcodes were pooled with an equivalent starting volume of the original sample (containing strain barcodes) on a sample-by-sample basis and processed together in downstream steps. Subsequently, a 2-cycle PCR protocol was performed to introduce unique molecular identifiers (UMIs) and dual sample indexes. Primer sequences were provided in Supplementary Table 4 (Primers 5 and 6). Two-sided size selection was performed using magnetic bead to sample ratios of 0.95X and 0.8X. The cleaned-up PCR products were then pooled by 96-well plate and amplified by PCR using 20 PCR cycles (Supplementary Table 4, Primers 7 and 8).

The final PCR product was run on a 2% agarose gel for clean-up and size selection. Zymoclean Gel DNA Recovery Kit (Ref #: D4002) was used to extract the DNA. After being denatured, diluted and mixed with a Phi-X spike-in of 30%, the DNA libraries were sequenced using 151 base-pair, paired-end reads using an Illumina MiniSeq (System suit version 2.2.1, Control software version 2.2.1). Data collection was achieved using the MiniSeq Mid Output Kit (300 cycles; Illumina Cat #: FC-420-1004). Paired end 151 base pair reads were performed with two 8 base pair index reads. The Local run manager (Version 2.4.1) was used to convert from .bcl to .fastq file formats using the GenerateFASTQ Analysis Module (V2.0.1).

NGS data analysis pipeline

The NGS data were processed following the previously published protocols⁵⁶. Galaxy (Galaxy version 20.05), an open-source and web-based platform, was used for NGS data analysis. After quality control using FastQC software (Galaxy Version .72+galaxy1)⁵⁹, pooled paired-end reads were demultiplexed using Barcode Splitter (Galaxy Version 1.0.1), Trim Sequences Galaxy (Version 1.0.2+galaxy0), and FASTQ joiner (Galaxy Version 2.0.1.1+galaxy0). One base-pair mismatch was allowed during the demultiplexing protocol. Sample indexes were 8 base pairs in length and had a minimum hamming distance of 3 base pairs. Paired-end reads were merged using the software FLASH (Galaxy Version 1.2.11.4)⁶⁰. Duplicate reads were removed using UMIs: Sort Collection (Galaxy Version 1.0.0), FASTQ joiner (Galaxy Version .0.1.1+galaxy0), UMI-Tools extract and deduplicate (Galaxy Version 0.5.5.1), Bowtie2 (Galaxy Version .3.4.3+galaxy0), and Samtools fastx (Galaxy Version 1.9+galaxy1). Strain specific barcodes were counted using a custom python script applying a matching algorithm. One mismatched base pair was allowed per barcode.

Calibration of strain barcodes and replicates

Samples for calibration were generated by mixing barcoded strains at three different known ratios, including one sample where all strains were mixed at equal volumes (Extended Data Fig. 6c). The strain compositions of calibration samples were then measured by NGS. The measured ratios of different strains correlated well with the expected ratios (Extended Data Fig. 6d), demonstrating the reliability of our barcoding approach.

Calibration of plasmid barcodes

Data correction was required in order to correlate NGS read counts of mobilizable plasmids (ESBL plasmids) to the relative abundance of plasmid-carrying cells in the population. Calibration samples were prepared using Keio strains each carrying one of the ESBL plasmids mixed at known ratios. One sample contained all Keio strains/plasmids mixed at equal volume. Since each Keio strain was carrying one of the mobilizable plasmids and all strains/plasmids were present at an equal ratio (i.e. the total plasmids abundance=1), we could define the relative abundance of each plasmid in total population as 1/ (total # plasmids). Using the known final abundance of each plasmid, a plasmid-specific correction factor was defined for each plasmid (denoted as $C_1$). These correction factors were subsequently applied to other calibration samples where plasmids were mixed at predefined ratios in order to verify the accuracy data correction. Plotting the expected sample relative abundance against the measured relative abundance for this calibration sample before and after data correction demonstrates the accuracy of this approach (Extended Data Fig. 6e).

Due to the fact that strain and plasmid barcodes underwent different sample preparation protocols (i.e. the addition pooled PCR protocol required for extraction and adaptor ligation of plasmid barcodes), we could not assume a 1:1 ratio between sequencing reads of strain and plasmid barcodes. Thus, a correction was required to account for the difference between the total number of NGS reads of strain-specific and plasmid-specific barcodes in order to define the relative abundance of each plasmid in the total population. This correction factor was defined using the calibration sample described above (containing Keio strains each carrying one mobilizable plasmid, mixed at equal ratio) diluted at known ratios with a mixture of the remaining barcoded Keio strains (not carrying any additional mobilizable plasmids) mixed at equal ratio. By diluting the mixture of strains containing an equal abundance of all mobilizable plasmids with additional barcoded Keio strains, we predictably varied the total relative plasmid abundance in the population. A correction factor (denoted as $C_2$) was defined to scale the number of plasmids reads such that the total plasmid abundance in the population accurately reflected the known total plasmid abundance when normalizing the number of NGS reads per plasmid by the total number of NGS counts in both plasmid and strain barcodes.

Taken both data corrections together, the relative abundance of each plasmid was calculated using the format below (example for plasmid A):

$$RelativeAbundanc{e}_{plasmidA}=\frac{\#Read{s}_{plasmidA}*C_2}{\#TotalNGSReads*C_1}$$

Where $C_1$ is the plasmids-specific correction, $C_2$ is the correction to correlate Keio and plasmid-specific NGS reads, and #TotalNGSReads is the sum of all NGS reads for plasmids and Keio strains per sample.

Model construction

To analyze the functional stability of complex communities, we considered a population of $m$ species transferring $n$ plasmids. To model the population dynamics, we applied a computational framework that we previously developed¹⁹, which is comprised of two groups of ordinary differential equations (ODEs):

$$\frac{d{s}_i}{dt}={\alpha }_i{\mu }_i^e{s}_i-D{s}_i,$$

$$\frac{d{p}_{ij}}{dt}={\beta }_{ij}{\mu }_{ij}^e{p}_{ij}+\left(s_i-p_{ij}\right)\sum _{k=1}^m{\eta }_{jki}{p}_{kj}-\left({\kappa }_{ij}+D\right)p_{ij}.$$

where $s_i$ is the abundance of the $i$-th species ($i=1,2,\dots ,m$) and $p_{ij}$ is the abundance of the $i$-th species carrying the $j$-th plasmid ($j=1,2,\dots ,n$). The first equation describes the growth dynamics of each species, where ${\mu }_i^e$ is the effective growth rate, and ${\alpha }_i$ represents the combined fitness costs of all the plasmids carried by the species. $D$ is the dilution rate. The second describes the plasmid dynamics. ${\mu }_{ij}^e$ is the effective growth rate of the $i$-th species carrying the $j$-th plasmid; it differs from ${\mu }_i^e$ due to the fitness effect of the j-th plasmid. ${\beta }_{ij}$ is the combined fitness costs of all the other plasmids. ${\eta }_{jki}$ is the transferring rate of the $j$-th plasmid from the $k$-th species to the $i$-th species. $s_i-p_{ij}$ is the abundance of the plasmid-free cells, which serve as the recipients of plasmid conjugation. ${\kappa }_{ij}$ is the plasmid loss rate. We also integrated the species interactions into the population dynamics via ${\mu }_i^e={\mu }_i^0\left(1-p_i/c_i\right)+\frac{{\delta }_i\sum {\gamma }_{ij}^{+}{p}_j}{\sum {\gamma }_{ij}^{+}{p}_j+1}-\sum {\gamma }_{ij}^{-}{p}_j$. Here, ${\mu }_i^0$ is the maximum growth rate of the $i$-th species, and $c_i$ is the maximum carrying capacity. ${\gamma }_{ij}^{+}$ and ${\gamma }_{ij}^{-}$ represent the strength of positive and negative interactions between the $i$-th species and the $j$-th species, respectively. ${\delta }_i$ describes the maximum growth benefit provided by all the positive interactions. This framework allows us to simulate the gene flow in complex communities.

In this framework, ${\mu }_i^e$ represents the effective growth rate of the ‘empty’ cells (cells that do not carry any plasmids), while ${\mu }_{ij}^e$ is the effective growth rate of the cells that are only equipped with the plasmid $j$. Let ${\lambda }_{ij}$ be the fitness costs of the $j$-th plasmid in the $i$-th species. The plasmid is burdensome with positive ${\lambda }_{ij}$ while beneficial with negative ${\lambda }_{ij}.$ The effective growth rate ${\mu }_{ij}^e$ is linked to ${\mu }_i^e$ via ${\lambda }_{ij}$. This framework assumes that the growth rate is inversely proportional to the fitness cost. Therefore, the relationship between ${\mu }_{ij}^e$ and ${\mu }_i^e$ can be obtained as ${\mu }_{ij}^e={\mu }_i^e/{(1+\lambda }_{ij})$. To obtain the combined cost of all the plasmids that $s_i$ carries, we calculated the weighted average of their costs as $\overline{{\lambda }_i}={\sum }_{j=1}^n\frac{p_{ij}}{s_i}{\lambda }_{ij}$. Then, the formulation of ${\alpha }_i$ can be obtained as ${\alpha }_i=\frac{1}{1+\overline{{\lambda }_i}}$, which leads to

$${\alpha }_i=\frac{s_i}{s_i+{\sum }_{j=1}^n(p_{ij}{\lambda }_{ij})}.$$

Similar to the definition of ${\alpha }_i,{\beta }_{ij}$ can be obtained as

$${\beta }_{ij}=\frac{s_i\left(1+{\lambda }_{ij}\right)}{s_i\left(1+{\lambda }_{ij}\right)+{\sum }_{\left\{k:k\ne j\right\}}(p_{ik}{\lambda }_{ik})}.$$

By formulating the parameters ${\alpha }_i,{\beta }_{ij},{\eta }_{jki}$ and ${\kappa }_{ij}$, the computational framework integrates different mechanisms underlying gene dynamics. For instance, ${\alpha }_i$ and ${\beta }_{ij}$ account for selective pressures, with positive selection and negative selection described by MGE fitness benefits and burdens, respectively. With given transfer rates ${\eta }_{jki}$, the framework can describe MGE flow in fully connected network or partially connected network and can be applied to non-mobilizable genes. By defining the initial distribution of MGEs in the species and their transfer routes, the framework can also take gene redundancy into account.

Numerical simulations of complex communities transferring multiple plasmids

To numerically simulate the functional stability of complex microbial communities, we first assembled 200 species pools, each was composed of 100 species transferring 20 plasmids. In each pool, we randomized all the kinetic parameters related to species growth and plasmids dynamics. Each parameter followed a uniform distribution in the specified range. Next, we allocated the species from the species pools to local populations by random sampling. For each pool, 40 local communities were assembled, and each local community contained 50 species on average. The dynamics of each local community was simulated for 500 hours until the community reached steady state, i.e. until the absolute abundances of each species and each plasmid remained unchanged. We calculated the steady-state abundance of plasmid in each local population, which then allowed us to calculate the abundance stability of each plasmid.

Extended Data

Extended Data Fig. 1|. Dynamic redundancy by HGT promoted the gene abundance stability of two-strain communities transferring a single plasmid.

a. The conjugation efficiency of R388 between MG1655 and Top10. Data are presented as mean values +/− standard deviations of three replicates. b. The growth curves of MG1655 and Top10 under different Strp concentrations. Data are presented as mean values +/− standard deviations of three replicates. c. The maximum growth rates of MG1655 and Top10 under different Strp concentrations. Data are presented as mean values +/− standard deviations of three replicates. d. The fitness burden of R388 in MG1655 or Top10. Data are presented as mean values +/− standard deviations of three replicates. e. Linoleic acid (LAC) inhibited the transfer of R388. The conjugations rates were normalized with the mean rate without LAC treatment. Data are presented as mean values +/− standard deviations of three replicates. f. Streptomycin did not impact the inhibition effects of linoleic acid on R388 transfer. The conjugations rates were normalized with the mean rate of the control group (0 mM LAC and 0 μg/mL Strp). Data are presented as mean values +/− standard deviations of three replicates. g. The temporal dynamics of Top10 relative abundances in the five communities during the experiment. Data are presented as mean values +/− standard deviations of three replicates. h. The relationship between community composition and R388 abundance at day 15. Data are presented as mean values of three replicates. i. The gene abundance stability $\varphi$ increases with the plasmid transfer rate in a model of two species transferring a single plasmid. The x-axis is divided into multiple bins with widths of 0.015. Error bars represent mean +/− standard deviation of the $\varphi$ values in each bin ($n=200$ independent data points).

Extended Data Fig. 2|. HGT promoted the gene stability of three-member communities transferring a single plasmid.

We assembled five communities transferring plasmid R388, each consisting of a P. aeruginosa strain (PA14) and two E. coli strains (MG1655 and Top10). The three members carried different antibiotic resistances (Cm, Tet and Strp, respectively), which allowed us to modulate the community composition by different antibiotic treatments. Specifically, the five communities (A to E) were treated with no antibiotic or with Cm, Tet, Strp or Strp+Tet, respectively. We further changed the plasmid transfer rate using LAC. a. The variations of the community compositions at the end of the experiment (day 15). The colored bars indicate different members, and the heights of the bars represent the relative abundances of the strains within the community. b. The dynamics of R388 abundance during the experiment. The results of three LAC concentrations and two dilution ratios were shown here. Data are presented as mean values +/− standard deviations of three replicates. c. The stability of plasmid abundance increased with the plasmid transfer rate. The results of day 5, 10 and 15 were shown in different line styles, while the two dilution ratios were represented by different marker styles.

Extended Data Fig. 3|. A two-species model predicts the plasmid R388 dynamics in monocultures or cocultures of MG1655 and Top10.

Dynamic parameters, including the plasmid transfer rates, strain growth rates under different Strp doses, plasmid burdens and LAC inhibition effects were measured and used to parameterize the model. a. Dynamics of R388 abundances in MG1655 or Top10 cultured seperately. For each strain, three LAC doses and two dilution ratios were tested. Experiment data are presented as mean values +/− standard deviations of three replicates. b. Dynamics of R388 abundances and strain compostions in cocultures of MG1655 or Top10. Five communities, treated by different concentrations of Strp, were modeled or experimentally measured. For each community, three LAC doses and two dilution ratios were tested. Experiment data are presented as mean values +/− standard deviations of three replicates. c. The degree of gene abundance stability in the MG1655-Top10 cocultures increased with the plasmid transfer rate. The results of day 5, 10 and 15 were shown in different line styles, while the two dilution ratios were represented by different marker styles.

Extended Data Fig. 4|. Numerical simulations demonstrated the stabilization of gene abundance mediated by HGT in complex communities.

a. A schematic of random sampling from a pool of interacting species to multiple local communities. The filled circles of different colors represent different species. The sizes of the circles describe the species abundances. The black arrows stand for positive inter-species interactions while the red arrows represent negative interactions. b. The stability of plasmid abundance against species fluctuations is promoted by HGT. Each column represents a single community. The colored bars stand for different species (first panel) or plasmids carrying functional genes (second to fourth panel), and the heights of the bars represent the relative abundances of species or plasmids within the community. The 40 communities differ from each other in their species compositions, due to random sampling. With slow plasmid transfer, the plasmid abundances vary drastically across different communities, while with rapid transfer, the functional profile is stable against compositional changes. Three different HGT rates (0.002, 0.005 and 0.01, from left to right) are shown as examples here. c. The gene stability $\varphi$ increases with the plasmid transfer rate. Here, 200 species pools are created, and each consists of 100 species transferring 20 plasmids. For each species pool, we randomize the parameters in the ranges of $-0.05<\gamma <0.05,0<\delta <0.2,0.2<\mu <0.6,0<\kappa <0.05$, $-0.05<\lambda <0.05$, and $0<\eta <0.1$. 40 local communities were created for every species pool by random sampling. On average, each local population contained 50 species. Then we simulated the community dynamics until it reached a steady state and calculated the $\varphi$ values of the plasmids as a function of their mean transfer rates. The x-axis is divided into multiple bins with widths of 0.005. Error bars represent mean +/− standard deviation of the $\varphi$ values in each bin ($n=400$ independent data points).

Extended Data Fig. 5|. The stability of plasmid abundance increases with the plasmid transfer rate even before the system has reached equilibrium.

Here, 200 species pools were assembled in silico, and each was composed of 100 species transferring 20 plasmids. For each master community, we randomized the parameters in the ranges of $-0.05<\gamma <0.05,0<\delta <0.2,0.2<\mu <0.6,0<\kappa <0.05$, $-0.05<\lambda <0.05$, and $0<\eta <0.1$. 40 local communities were created for every master community by random sampling. On average, each local population contained 50 species. The $\varphi$ values of the plasmids in the local communities were calculated by numerical simulations. The relationships between the stability $\varphi$ and mean plasmid transfer rate at 4 different timepoints ($t$ = 70, 100, 200 and 500 hours) were shown. Error bars represent mean +/− standard deviation of the $\varphi$ values in each bin ($n=400$ independent data points).

Extended Data Fig. 6|. Construction of barcoded Keio strains and sequencing quantification.

a. The backbone of barcoded plasmids. Structure of plasmid vector includes the origin of replication (p15A), a selection marker (chloramphenicol), a fluorescence marker (mCherry) and the unique barcode sequence. b. Sequence of the synthesized DNA fragments. Sequence of the synthetic barcode when assembled into vector backbone includes the 15-20 base-pair overlap with the plasmid vector used for Gibson Assembly (orange), Illumina^® adapter sequences (green) and two 18 base-pair barcode sequences (purple). c. Design of calibration experiments. Samples prepared at known concentrations were generated and underwent NGS library preparation, sequencing and data analysis pipelines. Images represent the layout of samples where each well contained a single barcoded strain. Strains were mixed together based on the indicated layout by column where darker shading indicates higher densities. Samples were prepared using a 2-fold dilution in each group. d. Overall calibration results. Normalized relative abundance of sequencing counts obtained for each barcode plotted versus the expected sample ratio shows a good correlation between expected and actual barcode abundances. Each data point represents a single barcoded Keio strain. e. Correction of NGS plasmid barcode reads. Results of correcting calibration samples using different data correction approaches. The measured relative abundance of NGS read counts for each plasmid barcodes was plotted against the expected relative abundance of samples with no data correction (left), empirical correction (middle graph) by normalizing reads with equal ratio NGS sample, and numerical correction (right graph) using plasmid-specific correction coefficient, respectively. R-squared values suggest an improved correlation between expected and actual relative abundance when applying numerical data correction.

Extended Data Fig. 7|. HGT stabilized gene abundances against compositonal variations in complex communities.

a. Different Keio strains transferred plasmid p13 with different conjugation rates. Here, Top10 carrying p13 served as the donor and 72 Keio strains served as recipients. Data are presented as mean values +/− standard deviations of three replicates. b. The temporal dynamics of strain compositions of Keio communities transferring plasmid p13 during the experiment. The results of three linoleic acid (LAC) concentrations and two dilution ratios were shown. In each panel, the colored bars stand for different Keio strains, and the heights of the bars represent their relative abundances within the community. c. The 13 plasmids were transferred within the synthetic communities with different conjugation efficiencies. Here, Top10 strain carrying the each of the 13 plasmids served as the donors, and Keio strain 1 served as the recipient. Data are presented as mean values +/− standard deviations of three replicates.

Extended Data Fig. 8|. The functional stability ϕ increases with the plasmid transfer rate even when different species express the genes at different levels.

x_ij (0 ≤ x_ij ≤ 1) describes the relative expression level of each copy of the $j$-th plasmid in the $i$-th species. The numerical simulations were performed following the same protocols as described in Extended Data Fig.4c and Methods of the main text. Four different ranges of $x_{ij}$ were considered, corresponding to different magnitudes of heterogeneity of gene expression levels. The x-axis is divided into multiple bins with widths of 0.005. Error bars represent mean +/− standard deviation of the $\varphi$ values in each bin ($n=400$ independent data points).

Supplementary Material

Source Data Extended Data Figure 3

Plasmid dynamics in monocultures of MG1655 and Top10

Source Data Extended Data Figure 2

Population dynamics of three-member communities transferring a single plasmid

Source Data Extended Data Figure 1

Kinetic parameters of two-strain communities

Source Data Extended Data Figure 6

Calibration of sequencing results

Source Data Extended Data Figure 7

Plasmid transfer rates and composition dynamics of Keio communities transferring thirteen plasmids

Source Data Figure 3

R388 abundances in two-strain communities

Source Data Figure 4

Population dynamics of Keio communities transferring a single plasmid

Source data Figure 5

Population dynamics of Keio communities transferring thirteen plasmids

Data availability:

Experimental data generated for this manuscript are deposited at GitHub at https://github.com/youlab/GeneStability_NCB2022. Source data are provided with this paper.