Data-driven prediction of colonization outcomes for complex microbial communities

Wu, Lu; Wang, Xu-Wen; Tao, Zining; Wang, Tong; Zuo, Wenlong; Zeng, Yu; Liu, Yang-Yu; Dai, Lei

doi:10.1038/s41467-024-46766-y

Download PDF

Article
Open access
Published: 16 March 2024

Data-driven prediction of colonization outcomes for complex microbial communities

Nature Communications volume 15, Article number: 2406 (2024) Cite this article

4616 Accesses
22 Altmetric
Metrics details

Subjects

Abstract

Microbial interactions can lead to different colonization outcomes of exogenous species, be they pathogenic or beneficial in nature. Predicting the colonization of exogenous species in complex communities remains a fundamental challenge in microbial ecology, mainly due to our limited knowledge of the diverse mechanisms governing microbial dynamics. Here, we propose a data-driven approach independent of any dynamics model to predict colonization outcomes of exogenous species from the baseline compositions of microbial communities. We systematically validate this approach using synthetic data, finding that machine learning models can predict not only the binary colonization outcome but also the post-invasion steady-state abundance of the invading species. Then we conduct colonization experiments for commensal gut bacteria species Enterococcus faecium and Akkermansia muciniphila in hundreds of human stool-derived in vitro microbial communities, confirming that the data-driven approaches can predict the colonization outcomes in experiments. Furthermore, we find that while most resident species are predicted to have a weak negative impact on the colonization of exogenous species, strongly interacting species could significantly alter the colonization outcomes, e.g., Enterococcus faecalis inhibits the invasion of E. faecium invasion. The presented results suggest that the data-driven approaches are powerful tools to inform the ecology and management of microbial communities.

Nutritional and host environments determine community ecology and keystone species in a synthetic gut bacterial community

Article Open access 08 August 2023

Top-down identification of keystone taxa in the microbiome

Article Open access 04 July 2023

Ecology-guided prediction of cross-feeding interactions in the human gut microbiome

Article Open access 26 February 2021

Introduction

Microbial communities are constantly exposed to the invasion of exogenous species, which can significantly alter their composition and function^1,2,3,4. The capacity of a microbial community to resist invasion is regarded as an emergent property (i.e., from individual parts to a holistic function) arising from the complex interactions among its constituent species⁵. Theoretical studies have found that communities with higher diversity or stronger interaction strengths among species are more resistant to potential invaders^6,7,8, attributed to the fact that communities with higher diversity can occupy more niches and provide functional redundancy, making it more difficult for an invading species to establish and thrive.

The role of host-associated microbiota in defending against pathogens has been extensively studied^9,10,11,12, particularly in the context of the human gut microbiome. For the human gut microbiome, invasion of resident microbial communities can occur when non-resident bacteria from foods and the upper gastrointestinal tract reach the gut ecosystem^13,14. The resident microbes in the gut ecosystem outcompete and exclude invaders through a combination of mechanisms, such as producing antimicrobial compounds^15,16, competing for nutrients and space^17,18,19,20, and modulating the host’s immune response^21,22. However, the composition of the human gut microbiota can vary significantly across individuals^23,24 and over time^25,26. Dietary shifts, medication, and other environmental factors can greatly alter the composition of the gut microbial community^27,28. These interpersonal and dynamic variations in the gut microbiome can lead to significantly different colonization outcomes, such as resistance to pathogens^29,30 and probiotics^31,32. For example, antibiotics treatment often leads to a loss of diversity and promotes the invasion of exogenous species^30,33. The ability to predict and alter the colonization outcomes (i.e., prevent the engraftment of pathogens and promote the engraftment of probiotics) is critical for personalized microbiota-based interventions in nutrition and medicine.

Despite the accumulating empirical studies, predicting the colonization outcomes in complex communities, such as the human gut microbiome, remains a fundamental challenge due to limited knowledge of interspecies interactions. For a meta-community of N species (N ranges from hundreds to thousands for the human gut microbiome), we would need to isolate and culture all these species (which is already a formidable task) and conduct a considerable number of experiments to map pairwise interactions, not to mention higher-order interactions. Thus, novel approaches are needed to study the ecology of highly complex communities.

The key question is: can we achieve system-level predictions for complex ecological systems without requiring detailed mechanistic information? Colonization outcome can be viewed as a mapping from the community structure of a complex ecological system (i.e., the pre-invasion community profile) to its function (i.e., the post-invasion abundance of the invading species). Recently, the application of data-driven models (machine learning and deep learning) has shown great promise in predicting the emergent properties of complex biomolecules, such as protein structure (mapping from protein sequence to structure)³⁴, promoter strength (mapping from DNA sequence to function)³⁵. While statistical models have been used to decipher microbial interactions in synthetic human gut microbial communities³⁶ and mouse gut microbiota³⁰, the use of data-driven models has not received much attention in microbial ecology^37,38.

Here, we proposed a data-driven approach to predict colonization outcomes of exogenous species in complex microbial communities. First, we systematically evaluated the approach using synthetic data generated by classical ecological dynamical models. We found that, with sufficient sample size in training data (on the order of ~O(N)), the colonization outcomes (i.e., whether an exogenous species can engraft and what its abundance would be if it can engraft) can be predicted by machine learning models. Then, we generated large-scale datasets with in vitro experimental outcomes of two representative species (E. faecium and A. muciniphila) colonizing human stool-derived microbial communities. We validated that machine-learning models, including Random Forest and neural ODE, can also predict colonization outcomes in experiments (AUROC > 0.8). Finally, we used the machine learning models to infer species with large colonization impacts and experimentally demonstrated that the introduction of strongly interacting species can significantly alter the colonization outcomes. Our results suggest that the colonization outcome of complex microbial communities can be predicted via data-driven approaches and tunable.

Results

The data-driven approach of predicting colonization outcomes for complex microbial communities

Let’s consider a meta-community with a pool of N microbial species, denoted as Ω = {1,⋯,N}. Consider a large set of M microbiome samples, denoted as S = {1,…,M}, collected from this meta-community. A microbiome sample s∈S can be considered as a local community of the meta-community with a subset of co-existing species (Fig. 1A). For a local community s, if an exogenous species i (still in the species pool Ω, but not in community s) is introduced to community s, whether it can successfully colonize the community or not, as well as its post-invasion abundance x_i, will depend on the baseline composition of community s. For example, it is easier (or harder) for species i to colonize community s if some resident species strongly promote (or inhibit) its growth of species i, respectively. Hereafter, we call the community s permissive (or resistant) to species i if species i can (or cannot) successfully colonize community s, respectively. If we only have the information about species i and a community s, it may seem impossible to accurately predict the colonization outcome without detailed knowledge about microbial interactions. However, if we have access to the data from colonization experiments of many local communities, then, in principle, we can formalize the colonization outcome prediction problem as a machine-learning task that can be solved in a data-driven fashion. To ensure the problem is mathematically well-defined, we must assume that the different local communities in this meta-community share identical assembly rules and microbial interactions³⁹. This way, the colonization outcomes of some local communities can be used to train a machine-learning model to predict the colonization outcomes of other local communities.

**Fig. 1: Prediction of colonization outcomes for complex microbial communities via the data-driven approach.**

Consider species-i as the exogenous species to a local community s. Note that the baseline abundance of species-i is zero (i.e., \({x}_{i}^{(0)}=0\)) in s before the invasion. With some initial abundance, the exogenous species will interact with the resident species in s, and its post-invasion steady-state abundance is denoted as \({x}_{i}^{(1)}\). We propose to solve the Colonization Outcome Prediction (COP) problem using machine-learning models that treat the baseline (i.e., pre-invasion) taxonomic profile \({{{{\bf{x}}}}}^{(0)}\) as inputs and the steady state abundance of the invasive species \({x}_{i}^{(1)}\) as output (Fig. 1B). Mathematically, we intend to learn the mapping from the baseline taxonomic profile of a community \({{{{\bf{x}}}}}^{(0)}\) to the steady state abundance of the invading species \({x}_{i}^{(1)}\), i.e., \(\varphi :{{{{\bf{x}}}}}^{(0)}\mapsto {x}_{i}^{(1)}\). In addition, this mapping could help us infer the impact of each resident species on the colonization of the exogenous species.

We conducted in silico simulations to validate the feasibility of our approach. We generated synthetic data of colonization outcomes using the Generalized Lotka–Volterra (GLV) model with N = 100 species in the meta-community (see Methods)⁴⁰. The initial species collection of each sample (i.e., a local community) consists of 30 species randomly drawn from the (N−1) species pool (the exogenous species is absent in all the local communities). We generated the baseline profiles of local communities by running the GLV dynamics to a steady state. The exogenous species was then added to each local community, and its post-invasion abundance was obtained by running the GLV dynamics to a new steady state.

We can formalize COP as two sub-problems: (1) Classification: predict whether an exogenous species can colonize a local community; (2) Regression: predict the steady-state abundance of an exogenous species after colonization. Using the synthetic data generated by the GLV model, we first addressed the classification problem, i.e., predicting whether the invading species can colonize a community. We employed three models covering representative categories of machine learning: Logistic Regression, Random Forest classifier, and COP-Neural Ordinary Differential Equations (COP-NODE) classifier (see Methods). We tuned the complexity of the ecological network (i.e., network connectivity) and evaluated the performance of different models at varying levels of the training sample size (Fig. 1C–E). Here, the network connectivity represents the probability of two species in the species pool interacting with each other. As expected, we observed that the predictive performance of machine learning models improved with the number of training samples. For network connectivity C = 0.3, we found that the Area Under the Receiver Operating Characteristic curve (AUROC, a perfect classifier has AUROC = 1 and AUROC = 0.5 for random guess) of three machine learning models was above 0.9 with training sample size \({S}_{{{{\rm{train}}}}}=N\). For higher network connectivity (e.g., C = \(0.4\) and \(0.5\)), the increased complexity in inter-species interactions rendered the binary prediction of colonization outcomes more difficult. Nevertheless, with a sample size on the order of ~O(N) per colonizing species, machine learning models were able to achieve accurate classification of colonization outcomes in synthetic data (AUROC > 0.8).

Next, we addressed the regression problem, i.e., predicting the steady-state abundance of the exogenous species. For the GLV model (with the interaction matrix A being invertible, which is almost surely true for randomly constructed matrices), our analytical derivations discovered a surprisingly simple linear relation between the post-invasion abundance of the exogenous species and the pre-invasion abundance of resident species (Supplementary Text, Fig. S1). Although the linear relation doesn’t hold for other dynamical models, it suggests that learning the mapping for COP is feasible by the data-driven approach, and the number of parameters required for fitting the relation is on the order of ~O(N). We employed three machine learning models: Elastic Net Linear Regression (ENET), Random Forest regressor, and COP-NODE regressor (Fig. 1F–H). The predictive performance was evaluated with Pearson’s correlation coefficient between the predicted and true abundance (log-transformed), as well as the ratio between the predicted abundance and the true abundance (Fig. S2). We systematically examined the predictive performance of three models at varying levels of network connectivity and training sample size. Similar to the classification problem, we found that increasing network connectivity \(C\) rendered the regression problem more difficult. For training sample size \({S}_{{{{\rm{train}}}}}=2N\) or higher, there was a substantial improvement in the quantitative prediction of the post-invasion abundance by ENET and NODE; in contrast, Random Forest had a poor performance at all sample sizes. Finally, we added varying levels of noise in the simulated data to assess the robustness of machine learning models to technical variations (e.g., measurement errors). For both the classification problem and the regression problem, we found that the predictive performance of machine learning models is robust against noise (Fig. S3).

Generation of human stool-derived in vitro microbial communities with diverse compositional profiles

To systematically study colonization outcomes in complex microbial communities, we used cultivation of human stool-derived in vitro communities in multi-well plates^41,42,43,44 (Fig. 2A, Methods). Briefly, we cultured gut microbial communities derived from 24 donors to reach steady states after five rounds of serial passaging in vitro. To increase the diversity in baseline communities, we treated each donor’s sample with a single pulse of 12 antibiotics from different classes (Table S1). After 24 h of antibiotics treatment, in vitro microbial communities were passaged every 24 h with a 1:200 dilution into fresh medium (Fig. 2A). Overall, we obtained more than 300 baseline communities with substantial variation in the compositional profiles at the species level (Fig. 2B, Figs. S4, S5). The compositional profiles of the baseline communities were stable, with around 40 to 120 co-existing species in each community (Fig. S6).

**Fig. 2: Invasion experiments in human stool-derived in vitro microbial communities.**

For the invasion experiments, we would introduce an exogenous species into the baseline communities and determine its colonization outcome after 8–10 rounds of serial passaging (Fig. 2C). We conducted a preliminary experiment to investigate the colonization outcome of different exogenous species (Fig. S7). We found that E. faecium, A. muciniphila, and Fusobacterium nucleatum could successfully colonize in some communities at varying levels of post-invasion abundance. In contrast, Streptococcus salivarius, Bifidobacterium breve, and Lactobacillus spp. could not colonize in nearly all the communities we tested. Moreover, vancomycin treatment significantly altered the colonization outcomes, rendering the gut microbial communities more susceptible to invasion (Fig. S7C). Overall, our results support the use of human stool-derived in vitro communities as a model experimental system for studying colonization outcomes.

Colonization outcomes of E. faecium in human stool-derived in vitro communities

We selected E. faecium as a representative species for colonization experiments in human stool-derived in vitro communities. E. faecium is a Gram-positive bacterium that inhabits the gut of humans and other animals. Some E. faecium strains have probiotic potential⁴⁵, and recent studies suggest that it plays a positive role in cancer immunotherapy^46,47. On the other hand, some E. faecium strains cause opportunistic infections in hospitalized patients with disrupted gut microbiota⁴⁸.

We introduced E. faecium to ~300 baseline communities (Fig. S8) at a dose of 5% relative to the total abundance of resident species. We passaged all communities for ten rounds to reach the post-invasion steady state (Methods). We observed that the colonization outcomes of E. faecium in different communities were persistent during serial passaging (Fig. S8B). In addition, the composition of in vitro communities before and after E. faecium invasion is highly reproducible across three replicates (Fig. S9).

We found that E. faecium was able to colonize 32% of baseline communities (i.e., permissive), with its post-invasion absolute abundances (estimated by multiplying its relative abundance with the total biomass OD₆₀₀) in permissive communities varying over two orders of magnitude (Fig. 3A). Previous studies suggested that community biomass and diversity are important factors underlying the colonization resistance to exogenous species^49,50. For example, reduced diversity of the resident community is often linked to pathogen infection in the human gut or other ecosystems⁵¹. Indeed, we found that the biomass and species richness of the baseline communities exhibited a clear negative correlation with the post-colonization abundances of E. faecium (Fig. S10). The diversity of the E. faecium permissive communities was significantly lower than the resistant communities (Fig. 3B, C). Furthermore, we observed a significant difference between the composition of E. faecium permissive communities and resistant communities (Fig. 3D). The colonization success of E. faecium was highly baseline-dependent, with substantial variations across different donors and antibiotics treatments (Fig. S11).

**Fig. 3: The colonization outcome of *E. faecium* in human stool-derived in vitro microbial communities is baseline-dependent and discriminable.**

For the regression problem, we need training samples with non-zero post-invasion abundance. Because E. faecium only colonized in ~30% of baseline communities in our experiments, the number of samples is insufficient to train the regression models to predict the post-invasion absolute abundance. To predict the binary colonization outcomes (permissive vs. resistant) of E. faecium, we employed three machine learning models, including Logistic Regression, COP-NODE classifier, and Random Forest classifier (Fig. 3E–G, Fig. S12AB). For 6-fold cross-validation, we used the communities derived from 20 donors (~240 samples) to train the model and the communities derived from the remaining 4 donors (~60 samples) to evaluate the model. Random Forest classifier displayed the best performance in predicting whether E. faecium could successfully colonize based on the species-level community composition (AUROC = 0.86, Accuracy = 0.82), followed by COP-NODE classifier (AUROC = 0.81, Accuracy = 0.81) and Logistic Regression (AUROC = 0.71, Accuracy = 0.75, and the Accuracy of a naive classifier that predicts E. faecium cannot colonize in all communities is around 0.7). We also evaluated the performance of machine learning classifiers with a balanced split of training and test samples, showing that Random Forest remains the best classifier (AUROC = 0.81 for E. faecium colonization, Fig. S13A). For comparison, we used the community diversity (quantified by the relative species richness) as the only feature to predict colonization outcome (see Fig. S14). Our results indicated that the relative species richness alone can be used as a predictor, but its prediction performance (with average AUROC = 0.78 and Accuracy = 0.32) is worse than elaborate classifiers, e.g., Random Forest using the taxonomic profile (with average AUROC = 0.86 and Accuracy = 0.82). Overall, our colonization experiments of E. faecium in complex human gut microbial communities validated that the data-driven approach can solve the classification problem of COP.

Quantitatively predict the colonization outcomes of A. muciniphila

To investigate the generality of our approach, we selected A. muciniphila as a second representative species for colonization experiments in human stool-derived in vitro communities. A. muciniphila is a Gram-negative mucin-degrading bacterium that inhabits the human gut. Due to its potential beneficial effects on human health^52,53,54, A. muciniphila is considered a promising probiotic candidate⁵⁵. A. muciniphila is found in the gut microbiome of around 30% of adults, and its abundance varies substantially across individuals⁵⁶. Similar to the experimental design of E. faecium, we introduced A. muciniphila to ~300 baseline communities at a dose of 5% relative to the total abundance of resident species and passaged all communities for eight rounds to reach the post-invasion steady state. The colonization outcome of A. muciniphila in different communities was persistent during serial passaging (Fig. S15), and the composition of in vitro communities post A. muciniphila invasion is highly reproducible across three replicates (Fig. S16).

Overall, we found substantial variations in the post-invasion steady-state abundance of A. muciniphila across different donors and antibiotics treatments (Fig. S17). A. muciniphila could colonize in 93.6% of baseline communities (i.e., permissive). For permissive communities, the post-invasion abundance of A. muciniphila displayed a bimodal distribution (Fig. 4A). We classified the permissive communities into two subgroups (high vs. low), depending on the post-invasion abundance of A. muciniphila (abundance threshold at \({10}^{-2}\)). The Shannon diversity (Fig. 4B) and species richness (Fig. 4C) of the A. muciniphila high permissive communities were significantly lower than those of the low permissive communities, and there was a significant difference between the community composition of the two groups (Fig. 4D).

**Fig. 4: The colonization outcomes of *A. muciniphila* in human stool-derived in vitro microbial communities are quantitatively discriminable.**

We evaluated the performance of machine learning models in predicting the colonization outcomes of A. muciniphila, both qualitatively (classification) and quantitatively (regression). Random Forest classifier displayed the best performance in binary classification (high permissive vs. low permissive of A. muciniphila) based on the species-level community composition (AUROC = 0.84), followed by COP-NODE classifier (AUROC = 0.79) and Logistic Regression (AUROC = 0.75) (Fig. 4E–G). To quantitatively predict the post-invasion abundance of A. muciniphila, we employed three machine learning models: ENET, COP-NODE regressor, and Random Forest regressor (Fig. 4H–J, Fig. S18). In comparison to the other two methods, the Random Forest regressor achieved the highest accuracy in quantitative prediction (Pearson’s correlation coefficient between the predicted and true abundances \({{{\rm{\rho }}}}=0.74,p < 2.2\times {10}^{-16}\)) and successfully recapitulated the bimodal distribution in the abundance of A. muciniphila. Taken together, we demonstrated the generality of the data-driven approach in predicting baseline-dependent colonization outcomes for complex microbial communities.

Colonization impact in simulated and experimental communities

Learning the mapping from the baseline taxonomic profile to colonization outcomes can help us infer the impact of each resident species on the colonization of the exogenous species (Fig.1B). To compute the colonization impact of regression (classification), we can perform a thought experiment by introducing a perturbation in the abundance of the resident species and use the trained machine learning model to predict the new colonization outcome of invading species after the perturbation (Fig. 5A). Negative colonization impact means that a resident species inhibits the colonization of the exogenous species in a given local community. In GLV simulations, while the colonization impact of resident species was randomly distributed (Fig. 5B), we found that the median colonization impact of a resident species across different local communities was positively correlated to its interaction strength on the exogenous species (Spearman correlation coefficient \(\rho=0.73,p < 2.2{\times 10}^{-16}\), Fig. 5C), suggesting that we may use colonization impact to identify strongly interacting species.

We used the Random Forest model to evaluate the colonization impact of all species in human stool-derived in vitro communities on E. faecium and A. muciniphila (Fig. 5D–G). We inferred that most resident species had a weak negative colonization impact (Fig. 5D, F). Based on the median colonization impact of a certain resident species across different local communities, we identified the top-ranking species with negative colonization impact (Fig. 5E, G). Colonization impact on A. muciniphila was overall less negative than E. faecium, consistent with our observation that human gut microbial communities were more permissive to A. muciniphila colonization.

The impact of strongly interacting species on colonization outcomes

To understand the role of strongly interacting species on colonization outcomes, we systematically studied the impact of E. faecalis on the colonization of E. faecium. E. faecalis was inferred to have the strongest colonization impact on E. faecium across different baseline communities (Fig. 5E). Besides, we found that E. faecalis, as well as other 5 species among the top 20 list (including Faecalibacterium prausnitzii, Ruminococcus gnavus, Blautia sp., Clostridium sp. L.2.50, and Roseburia inulinivorans) were predicted by both Random Forest (Fig. 5E) and NODE (Fig. S19AB) to have a strong impact on E. faecium colonization. In contrast, Logistic Regression classifier did not identify E. faecalis as a strong inhibitor (Fig. S19CD), and its predictive performance is substantially worse than Random Forest and NODE (Fig. 3). The colonization impact of those top-ranking species is also consistent with the result of LIME which is a novel explanation technique that explains the predictions of any classifier⁵⁷, e.g., the presence of E. faecalis, F. prausnitzii, and Clostridium perfringens in resistant communities increases the probability that E. feacium cannot colonize (see Fig. S19E).

We observed a statistically significant negative correlation between the abundance of E. faecalis in baseline communities and the post-invasion abundance of E. faecium (Kendall’s \({{{\rm{\tau }}}}=-0.37\), \(p=5.29{\times 10}^{-16}\)). In particular, baseline communities derived from some donors (e.g., S10, S07) had a high abundance of E. faecalis and were resistant to E. faecium colonization (Fig. 6A).

Fig. 6: The presence of *E. faecalis* in baseline communities inhibits the invasion of *E. faecium.*

We found that E. faecalis inhibited the growth of E. faecium in pairwise co-culture, either in liquid culture or on agar plates (Fig. S20). Then, we introduced E. faecalis into eight human stool-derived in vitro communities that were permissive to E. faecium invasion, using three different types of interventions (Fig. 6B–D, Fig. S21A): (1) add E. faecalis into the baseline community, followed by E. faecium on the next day; (2) add E. faecalis and E. faecium on the same day; (3) add E. faecium into the baseline community, followed by E. faecalis on the next day. In the control group, we only added E. faecium. In all three intervention groups, the colonization of E. faecium was significantly inhibited by E. faecalis across different baseline communities. Also, the inhibitory effect was consistent for two different E. faecalis strains isolated from human stool samples (Methods). In comparison, Clostridium symbiosum, a species predicted to have a neutral impact, did not alter the colonization of E. faecium (Fig. S21B).

Finally, we explored if the strong inhibition of E. faecalis on E. faecium could be shaping their distribution in the human gut via priority effects, i.e., the gut microbiome colonized with E. faecalis becomes resistant to E. faecium. We performed metagenomic sequencing of ~120 healthy volunteers in the SIAT cohort (Methods), whose samples were used to derive the in vitro communities and isolate the Enterococcus strains in this study. Indeed, there was a statistically significant negative correlation between the relative abundance of E. faecalis and E. faecium in the SIAT cohort (Kendall correlation \({{{\rm{\tau }}}}=-0.36,\,p=0.0044\), Fig. S22A). A similar pattern was observed in gut metagenomic samples of four independent cohorts (Kendall correlation τ = −0.36, p = 5.439 × 10⁻¹⁵, Fig. S22B).

Overall, our experimental validations and analysis suggest that data-driven models can infer species with strong colonization impact and guide the modulation of resident communities to alter the colonization outcomes of exogenous species.

Discussion

Here we proposed and systematically validated a data-driven approach to predict colonization outcomes of exogenous species, providing a powerful tool to inform the management of complex ecosystems. Pairwise co-culture^50,58,59,60 and synthetic communities^17,36,61,62 have been widely used to study the ecology and function of microbial communities. These experiments require the isolation and cultivation of individual species, thus are often limited to simple communities. In comparison, our approach is based on sampling an ensemble of complex communities (~100 species, Fig. S4) and using the sampled communities to infer the mapping between community composition and colonization outcomes⁶³ by assuming the compositional profiles represent steady states of the local communities. We demonstrate that the data-driven approach enables accurate function prediction and system-level understanding of complex microbial communities.

Understanding the colonization resistance of complex communities is a fundamental question in ecology. In our invasion experiments (~300 local communities and two different exogenous species), we found that resistance to exogenous species was positively correlated to community diversity, supporting the view that colonization resistance is an emergent property of complex communities⁸. Community diversity (i.e. species richness) alone can provide a reasonable degree of predictive accuracy (Fig. S14). However, it is less effective compared to utilizing full taxonomic profiles. While most resident species had a weak negative impact on the colonization of exogenous species, we identified E. faecalis as a strong inhibitor of E. faecium. We validated that introducing strongly interacting species into baseline communities can alter the colonization outcomes. It should be noted that the colonization impact is dependent on the community context (Fig. 5), because it takes into account both direct and indirect effects on the invading species⁶⁴, as well as potential higher-order interactions^65,66. Previous studies have shown that strongly interacting species can lead to priority effects⁶⁷, with important implications for community assembly in the infant gut microbiome and the formation of community types^68,69. Moreover, strongly interacting species can be used to modulate the resident communities to prevent the colonization of pathogens¹⁶ or facilitate the colonization of beneficial microbes (e.g., probiotics, crop fertilizers)⁷⁰.

Our results suggest that the colonization resistance of microbial communities is predictable and tunable via the data-driven approach, given that training data size is sufficient (on the order of ~O(N))⁷¹. For synthetic data generated by the classical GLV model in community ecology, we did see that the simple models, e.g., linear regression and ENET work well for both classification problems (Fig. 1C–E) and regression problems (Fig. 1F–H). However, those simple statistical models did not work well in predicting the colonization of A. muciniphila (Fig. 4H). We anticipate that real microbial communities are more complicated than the simple GLV model (which only includes pair-wise inter-species interactions). Sophisticated machine learning models may have to be leveraged to predict colonization outcomes for complex communities. We anticipate that more training samples are required if high-order interactions are considered. However, those high-order interactions might be weak and do not significantly impact the prediction, as the community-function landscapes display a low degree of ruggedness⁷². The high-throughput cultivation of gut microbial communities in vitro provides a powerful approach to studying the human gut microbiome^44,73. In our experiments, the number of species in the meta-community was ~160, and we profiled ~300 baseline communities for proof-of-concept validation. Meeting the sample size requirement for gnotobiotic plants is feasible^74,75. However, it could be challenging to gather sufficient training data for gnotobiotic animals and human cohort studies, depending on the complexity of the meta-community. In addition to data size, another critical concern is the technical variability in large-scale experiments⁷⁶. In future studies, experimental workflows can be automated to minimize technical variability and ensure data quality for training machine learning models.

Our study has several limitations. First, we did not account for potential variations at the strain level⁷⁷. Previous studies have shown that the strength of interspecies interactions can vary across different strains, such as the inhibition of Klebsiella pneumoniae by Klebsiella oxytoca^16,78. We also observed strain-level variations in the inhibition of E. faecium by E. faecalis (Fig. 6), and the underlying mechanism remains to be elucidated. Second, our invasion experiments in vitro did not reflect host-mediated interactions, which also contribute to colonization resistance in vivo²². Nevertheless, the higher permissiveness to A. muciniphila than E. faecium in human gut microbial communities in vitro is consistent with the higher prevalence of A. muciniphila in metagenomic samples^56,79. Third, we assumed that there was a single post-invasion steady state in simulated and experimental communities. The colonization outcomes may be influenced by multi-stability in microbial communities, e.g., successful colonization depends on the initial abundance of the invading species^80,81. Fourth, in our in vitro experiments, we found that A. muciniphila was able to stably colonize in the majority of stool-derived communities with relatively high abundance. It is known that mucin is the preferred nutrient source of A. muciniphila⁸², so it would be interesting to study to which degree the colonization of A. muciniphila depends on the mucin concentration provided in the medium. Lastly, while the data-driven framework can be generalized to different scenarios, the machine learning models must be re-trained when the environmental condition changes. In contrast, mechanism-based models can better deal with changes in conditions (e.g., pH, nutrient level).

We noted a potential difference between the GLV simulated data and experimental data: the exogenous species may already be present in the stool-derived communities, but its steady-state abundance was below the detection threshold. In this scenario, the introduction of the exogenous species (~5% of community biomass) may provide a growth boost (e.g., via some density-dependent mechanism) and enable the species to co-exist with other species at a higher steady-state abundance (i.e., multiple stable states). Moreover, there is a discrepancy between the performance of Random Forest in gGLV simulated data and real data. Potential explanations include: (1) the dynamics of the GLV model may be different from that of experimental communities. For instance, when the GLV model has globally stable equilibria, the final state is solely determined by the species collection. (2) the distribution of interspecies interaction strength used in the GLV model may differ from experimental communities. In experimental communities, a few strongly interacting species may dominate the contribution to the colonization resistance of exogenous species. In contrast, in simulated data, the contribution is more evenly distributed among resident species.

Our data-driven approach is independent of any dynamics model to predict colonization outcomes of exogenous species for complex microbial communities without detailed knowledge of the underlying ecological and biochemical process. We anticipate that the data-driven approach can be generalized to predict and engineer the function of microbial communities (i.e., mapping from community composition to function)^37,71,83,84. Similarly, this approach can be used to predict the response of microbial communities to various types of perturbations (i.e., mapping from community composition to the shift in composition/function), such as the baseline-dependent response of the human gut microbiome to prebiotics, food additives, etc., refs. ^85,86. In parallel to the breakthroughs in predicting the properties of complex biomolecules, we envision that the data-driven approach will lead to a paradigm shift in studying the stability and function of complex ecological systems and guide important applications in healthcare (e.g., personalized nutrition based on the human gut microbiome) and agriculture.

Methods

Collection and preservation of human stool samples

All human participants at SIAT (referred to as “SIAT cohort”) signed the informed consent form in the present study which was approved by the Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences (SIAT-IRB-200315-H0438). Stool samples were collected from healthy human donors and were immediately transferred to an anaerobic workstation (85% N₂, 10% H₂ and 5% CO₂, COY). 10 g of each stool sample was suspended into 50 mL 20% glycerol (v/v, in sterile phosphate-buffered saline, with 0.1% L-cysteine hydrochloride), homogenized by vortexing, and then filtered with sterile nylon mesh to remove large particles in fecal matter. Aliquots of the suspension were stored in sterile cryogenic vials and frozen at −80 °C for long-term storage until processing for DNA extraction and culturing so that the stool-derived community could be revived (thawed) for repeatable experiments.

Cultivation of human stool-derived in vitro communities

20ul stool slurries aliquot stocks were inoculated into 980 μL medium containing antibiotics in triplicate into 96 deep-well plates (PCR-96-SG-C, Axygen) for static culturing at 37 °C for 24 h in the anaerobic workstation. The concentration for each antibiotic was evaluated as described in the SI method. The medium (MiPro) used for in vitro culture was modified from previous studies, which comprises: peptone water (2.0 g /L, CM0009, Thermo Fisher), yeast extract (2.0 g /L, LP0021B, Thermo Fisher), L-cysteine hydrochloride (1 g/ L), Tween 80 (2 mL/L), hemin (5 mg/L), vitamin K1(10 μL/L), NaCl (1.0 g /L), K₂HPO₄ (0.4 g/L), KH₂PO₄ (0.4 g/L), MgSO₄⋅7H₂O (0.1 g/L), CaCl₂⋅2H₂O (0.1 g/L), NaHCO₃ (4 g/L), porcine gastric mucin (4 g/L, M2378, Sigma-Aldrich), sodium cholate (0.25 g/L) and sodium chenodeoxycholate (0.25 g/L)⁸⁷. After 24 h of antibiotics treatment, in vitro microbial communities were passaged every 24 h with a 1:200 dilution into fresh medium using the automated 96-format Thermo Scientific™ ClipTip™ (Thermofisher) pipette (every 24 h, 5 μL of this saturated culture was transferred into 995 μL of fresh medium). After 5 days of passaging, 500 μL of the cultures were mixed with 500 μL sterile 40% glycerol (v/v, in sterile phosphate-buffered saline, with 0.1% L-cysteine hydrochloride) in crimp vials, sealed, and stored as baseline communities at −80 °C for further usage and long-term storage. After each transfer, the remaining samples were centrifuged to remove the supernatant, and the pellets were stored at −80 °C with a plastic seal until DNA extraction. The in vitro microbial community biomass was evaluated by measurement of optical density (OD₆₀₀) with an Epoch 2 plate reader (BioTek) after 24 h of incubation.

Generation of baseline communities with diverse taxonomic profiles

To examine if in vitro stool-derived communities can reach stable states and display diverse compositions, we collected stool samples from healthy donors and grew them in MiPro medium, which has shown its capability in capturing and maintaining the diversity of in vitro stool-derived communities^42,87,88. We inoculated the stool aliquots into 96-well plates with growth media and incubated them in an anaerobic workstation in triplicate, passing them every 24 h with a 1:200 dilution. The microbial communities were assessed by shallow metagenomic sequencing, which is a cost-effective method for characterizing species-level composition of microbiota samples⁸⁹. We collected time-series data to examine the dynamics of community establishment on the in vitro platform. The metagenomic analysis revealed that, after an initial period of approximately four days, the composition profiles of almost all in vitro communities reached a stable and reproducible steady state. Our analysis also showed that the stool-derived in vitro communities were highly complex in their compositions and could retain personalized gut microbiota variation, as evidenced by species-level time-series compositions of 4 representative communities derived from 4 donors over ten rounds of in vitro passaging in MiPro (Fig. S6A, B).

From the fecal samples of SIAT cohort, we selected 24 donors in which E. faecium and A. muciniphila were not detected by metagenomic sequencing. To increase the diversity in baseline communities, we treated each donor’s sample with 12 antibiotics from different classes⁹⁰ (Fig. S4). Those stool-derived communities were treated with antibiotics for 24 h on Day 0 (i.e. a single pulse). Afterwards, the communities were passaged five times (from Day 1 to Day 6) in antibiotic-free medium to reach a stable state before introducing the exogenous species. Different antibiotic classes target distinct spectra of bacteria, leading to a remodeling of the community in different directions⁹⁰. We selected antibiotics from different classes as described in the EUCAST databases⁹¹. The optimal concentrations of the antibiotics were determined based on a previous study that evaluated the activity spectrum of antibiotic classes on human gut commensals⁹⁰. We tested at least three different concentrations for each antibiotic and evaluated the optimized dose based on its ability to partially inhibit (50–80%) the overall growth of stool-derived bacteria as measured by OD₆₀₀ after 24 h of incubation. To ensure reproducibility, we screened at least three different stool aliquot stocks as biological duplicates for each antibiotic. We measured the OD₆₀₀ of each well every 30 min using an Epoch 2 plate reader (BioTek) and collected growth curves up to 24 h.

Bacterial strains

Enterococcus faecium, Enterococcus faecalis, Clostridium symbiosum, Streptococcus salivarius, and Bifidobacterium breve strains were isolated from fecal samples of SIAT cohort. Taxonomy of isolates from SIAT cohort was confirmed by whole genome sequencing. Genome sequences have been deposited in PRJEB60398 (see “Data availability”). Lactobacillus plantarum HNU082⁹², Lactobacillus paracasei HNU312⁹³ was provided by Prof. Jiachao Zhang from Hainan University. Akkermansia muciniphila (ATCC BAA-835) and Fusobacterium nucleatum (ATCC 25586) were purchased from ATCC.

Profiling the colonization outcomes of different exogenous species

We conducted a preliminary experiment to investigate the colonization outcome of gut microbial communities to different exogenous species (Fig. S7), including: E. faecium, A. muciniphila⁹⁴, F. nucleatum, S. salivarius, B. breve, and Lactobacillus spp. (L. plantarum HNU082 and L. paracasei HNU312). We identified 12 stool samples from healthy donors in which the selected invader species were undetectable in the microbiota. We then cultured the stool samples in vitro and exposed them to antibiotics before introducing the exogenous species (~5% of total biomass, approximately 10⁶ CFUs for each well) into the community. We used shallow metagenomic sequencing to monitor the time-series and final community composition.

Invasion experiments of E. faecium and A. muciniphila

To conduct invasion experiments, frozen stocks of E. faecium (strain SIAT_DA797) and A. muciniphila (strain ATCC BAA-835) were grown anaerobically in BHI and mGAM at 37 °C, respectively, until stationary phase. In vitro microbial baseline communities, stored at −80 °C, were thawed and revived by adding 20 μL of the stocks to 980 μL of MiPro medium in deep-well plates. After incubation for 24 h at 37 °C, community biomass was measured by OD₆₀₀, and 5 μL of the saturated cultures were diluted into 1 mL of fresh MiPro in a new plate. Each well was invaded with the respective amount of E. faecium or A. muciniphila, with biomass representing 5% of the inoculated communities’ average biomass. The inoculum was passaged every 24 h of incubation, with a 1:200 dilution into fresh medium for 8–10 passages until the community reached a steady state (10 passages for E. faecium, 8 passages for A. muciniphila, based on data from Fig. S6). After each passage, the remaining samples were centrifuged to remove the supernatant, and the pellets were stored at −80 °C with a plastic seal in plate until DNA extraction.

Metagenomic sequencing and taxonomic profiling

DNA was extracted from 200 mg of stool samples using the QIAamp Power Fecal Pro DNA Kit (Qiagen) according to the manufacturer’s instructions. For stool-derived in vitro-cultured samples, 500 uL of cultured samples were used for DNA extraction with the DNeasy UltraClean 96 Microbial Kit (Qiagen) using an automated protocol at Tecan Freedom EVO 200. The Hieff NGS® OnePot II DNA Library Prep Kit for Illumina® (Yeasen) was used for library preparation, following the manufacturer’s instructions. The resulting library DNA was cleaned up and size-selected with Hieff NGS® DNA Selection Beads (Yeasen), and quantified using the dsDNA High Sensitivity kit on a Qubit (Thermo Fisher). Libraries were further pooled together at equal molar ratios, and the purity and library length distribution were assessed using Bioanalyzer High Sensitivity DNA Kit (Agilent). Sequencing was performed on the Illumina HiSeq X Ten system (150 bp paired-end reads; Annoroad Gene Technology Co.), with a target sequencing depth of 0.3 Gbp raw data per sample, as recommended by previous studies⁸⁹.

Samples with fewer than 10⁵ clean reads were excluded from downstream analysis. Prior to analysis, reads were trimmed using the following criteria: (1) Removing reads with more than 50% of the base below quality score 19; (2) Removing reads with more than 5% of the base being N; (3) Discarding paired-end reads if either of the paired reads did not meet the above criteria. Microbial community composition from metagenomic sequencing data was generated using the SHOGUN pipeline and the RefSeq database version 82, as described in previous studies^89,95. Species-level abundance profiles were filtered by using a relative abundance threshold of 0.0001 (0.001) for all taxa in colonization prediction of E. faecium (A. muciniphila), and those low-prevalence taxa (present in less than 20% samples) were further filtered to reduce the feature number. The colonization outcomes were evaluated based on the invader’s absolute abundance in the community, which was estimated by multiplying the relative abundance and the OD₆₀₀ value (OD₆₀₀ × relative abundance). To ensure repeatability, samples with Pearson correlation below 0.8 among replicates were excluded from COP analysis. This resulted in the exclusion of 1.8% of samples for E. faecium and 1.3% for A. muciniphila.

Quantification of the relative abundance of E. faecium and A. muciniphila by metagenomic sequencing

To confirm the accuracy of shallow metagenomic sequencing in quantifying the relative abundance of E. faecium and A. muciniphila, a spike-in experiment was conducted (Fig. S23A). In this experiment, a predefined amount of bacterial DNA from the target species was added to a metaDNA sample extracted from an in vitro community derived from human stool. This metaDNA sample was used as the background, since it has been previously sequenced and did not contain the target species. The spike-in DNA of the target species (E. faecium or A. muciniphila) was 1:10 diluted for eight times and was added to the microbial metaDNA to a mixed DNA sample (5 μL of target species DNA into 30 ng of microbial metaDNA). Three replicates were made for each sample. The mixed DNA was then used for library construction and metagenomic sequencing. By comparing the detected relative abundance generated by shallow metagenomic sequencing with the expected abundance, the accuracy and sensitivity of our workflow were determined. The detection threshold of E. faecium is 0.0001 (Fig. S23B) and the detection threshold of A. muciniphila is 0.001 (Fig. S23C). Our results showed that the quantification of the relative abundance of the two target species using the shallow metagenomic sequencing pipeline is accurate and reproducible.

Statistical analysis

Statistical details for each experiment are indicated in the figure legends. Pearson correlation coefficients and the p-values for testing replicates communities’ composition correlation were calculated on log₁₀(relative abundance). Kendall correlation coefficients and the p-values for testing E. faecium and E. faecalis abundance correlation were calculated on log₁₀(relative abundance). Alpha diversity of the community was calculated on species profile using the observed species richness and Shannon index. The composition of microbiota and variations in colonization outcomes between communities were analyzed by performing PCoA using the Bray-Curtis dissimilarity metric on the species-level abundance profile. Similarities among groups were determined by permutational multivariate analysis of variance (PERMANOVA, Adonis test) based on the Bray-Curtis dissimilarity⁹⁶, with 999 permutations used to test the significance. These analyses were conducted using the vegan⁹⁷ package (version 2.6–4). Non-parametric Mann–Whitney U-test were used to conduct pairwise comparisons between two groups⁹⁸. P values of less than 0.05 were considered as statistically significant, as indicated in the figures (ns, not significant, *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001). Data analysis and plotting was performed in R version 4.1.2 and R studio version 2022.12.0 + 353 using the packages dplyr, ggpubr, vegen, and ComplexHeatmap.

Reporting summary

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

Data availability

All sequencing data generated in this study are available from European Nucleotide Archive (ENA) under study accession number PRJEB60398. Sample accession code, metadata and related source data are provided as a Source data file with this paper. Source data are provided with this paper.

Code availability

The code for simulations and data analysis is available at https://github.com/spxuw/COP.

References

Buffie, C. G. & Pamer, E. G. Microbiota-mediated colonization resistance against intestinal pathogens. Nat. Rev. Immunol. 13, 790–801 (2013).
Article CAS PubMed PubMed Central Google Scholar
Gensollen, T., Iyer, S. S., Kasper, D. L. & Blumberg, R. S. How colonization by microbiota in early life shapes the immune system. Science 352, 539–544 (2016).
Article CAS PubMed PubMed Central ADS Google Scholar
Walter, J., Maldonado-Gomez, M. X. & Martinez, I. To engraft or not to engraft: an ecological framework for gut microbiome modulation with live microbes. Curr. Opin. Biotechnol. 49, 129–139 (2018).
Article CAS PubMed Google Scholar
Amor, D. R., Ratzke, C. & Gore, J. Transient invaders can induce shifts between alternative stable states of microbial communities. Sci. Adv. 6, eaay8676 (2020).
Article CAS PubMed PubMed Central ADS Google Scholar
van den Berg, N. I. et al. Ecological modelling approaches for predicting emergent properties in microbial communities. Nat. Ecol. Evol. 6, 855–865 (2022).
Article PubMed PubMed Central Google Scholar
Kennedy, T. A. et al. Biodiversity as a barrier to ecological invasion. Nature 417, 636–638 (2002).
Article CAS PubMed ADS Google Scholar
Kurkjian, H. M., Akbari, M. J. & Momeni, B. The impact of interactions on invasion and colonization resistance in microbial communities. PLoS Comput. Biol. 17, e1008643 (2021).
Article CAS PubMed PubMed Central ADS Google Scholar
Case, T. J. Invasion resistance arises in strongly interacting species-rich model competition communities. Proc. Natl Acad. Sci. USA 87, 9610–9614 (1990).
Article CAS PubMed PubMed Central ADS Google Scholar
Vonaesch, P., Anderson, M. & Sansonetti, P. J. Pathogens, microbiome and the host: emergence of the ecological Koch’s postulates. FEMS Microbiol. Rev. 42, 273–292 (2018).
Article CAS PubMed Google Scholar
Pamer, E. G. Resurrecting the intestinal microbiota to combat antibiotic-resistant pathogens. Science 352, 535–538 (2016).
Article CAS PubMed PubMed Central ADS Google Scholar
Li, P. D. et al. The phyllosphere microbiome shifts toward combating melanose pathogen. Microbiome 10, 56 (2022).
Article CAS PubMed PubMed Central Google Scholar
Mendes, R., Garbeva, P. & Raaijmakers, J. M. The rhizosphere microbiome: significance of plant beneficial, plant pathogenic, and human pathogenic microorganisms. FEMS Microbiol. Rev. 37, 634–663 (2013).
Article CAS PubMed Google Scholar
Schmidt, T. S. et al. Extensive transmission of microbes along the gastrointestinal tract. eLife 8, e42693 (2019).
Article PubMed PubMed Central Google Scholar
Kok, C. R. & Hutkins, R. Yogurt and other fermented foods as sources of health-promoting bacteria. Nutr. Rev. 76, 4–15 (2018).
Article PubMed Google Scholar
Chassaing, B. & Cascales, E. Antibacterial weapons: targeted destruction in the microbiota. Trends Microbiol. 26, 329–338 (2018).
Article CAS PubMed Google Scholar
Kim, S. G. et al. Microbiota-derived lantibiotic restores resistance against vancomycin-resistant Enterococcus. Nature 572, 665–669 (2019).
Article CAS PubMed PubMed Central ADS Google Scholar
Eberl, C. et al. E. coli enhance colonization resistance against Salmonella Typhimurium by competing for galactitol, a context-dependent limiting carbon source. Cell Host Microbe 29, 1680–1692 e1687 (2021).
Article CAS PubMed Google Scholar
Oliveira, R. A. et al. Klebsiella michiganensis transmission enhances resistance to Enterobacteriaceae gut invasion by nutrition competition. Nat. Microbiol. 5, 630–641 (2020).
Article CAS PubMed Google Scholar
Litvak, Y. et al. Commensal Enterobacteriaceae protect against Salmonella colonization through oxygen competition. Cell Host Microbe 25, 128–139.e125 (2019).
Article CAS PubMed Google Scholar
Lloyd, D. P. & Allen, R. J. Competition for space during bacterial colonization of a surface. J. R. Soc. Interface 12, 0608 (2015).
Article PubMed Google Scholar
Ivanov, I. I. et al. Induction of intestinal Th17 cells by segmented filamentous bacteria. Cell 139, 485–498 (2009).
Article CAS PubMed PubMed Central Google Scholar
Stecher, B. et al. Salmonella enterica serovar typhimurium exploits inflammation to compete with the intestinal microbiota. PLoS Biol. 5, 2177–2189 (2007).
Article CAS PubMed Google Scholar
Human Microbiome Project, C. Structure, function and diversity of the healthy human microbiome. Nature 486, 207–214 (2012).
Article ADS Google Scholar
Franzosa, E. A. et al. Identifying personal microbiomes using metagenomic codes. Proc. Natl Acad. Sci. USA 112, E2930–E2938 (2015).
Article CAS PubMed PubMed Central Google Scholar
Han, N. et al. Time-scale analysis of the long-term variability of human gut microbiota characteristics in Chinese individuals. Commun. Biol. 5, 1414 (2022).
Article PubMed PubMed Central Google Scholar
O’Toole, P. W. & Jeffery, I. B. Gut microbiota and aging. Science 350, 1214–1215 (2015).
Article PubMed ADS Google Scholar
Turnbaugh, P. J., Backhed, F., Fulton, L. & Gordon, J. I. Diet-induced obesity is linked to marked but reversible alterations in the mouse distal gut microbiome. Cell Host Microbe 3, 213–223 (2008).
Article CAS PubMed PubMed Central Google Scholar
Maier, L. et al. Extensive impact of non-antibiotic drugs on human gut bacteria. Nature 555, 623 (2018).
Article CAS PubMed PubMed Central ADS Google Scholar
Alavi, S. et al. Interpersonal gut microbiome variation drives susceptibility and resistance to cholera infection. Cell 181, 1533 (2020).
Article CAS PubMed PubMed Central Google Scholar
Buffie, C. G. et al. Precision microbiome reconstitution restores bile acid mediated resistance to Clostridium difficile. Nature 517, 205–U207 (2015).
Article CAS PubMed ADS Google Scholar
Maldonado-Gomez, M. X. et al. Stable engraftment of Bifidobacterium longum AH1206 in the human gut depends on individualized features of the resident microbiome. Cell Host Microbe 20, 515–526 (2016).
Article CAS PubMed Google Scholar
Zmora, N. et al. Personalized gut mucosal colonization resistance to empiric probiotics is associated with unique host and microbiome features. Cell 174, 1388–1405 e1321 (2018).
Article CAS PubMed Google Scholar
Liao, C. et al. Compilation of longitudinal microbiota data and hospitalome from hematopoietic cell transplantation patients. Sci. Data 8, 71 (2021).
Article PubMed PubMed Central Google Scholar
Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).
Article CAS PubMed PubMed Central ADS Google Scholar
Vaishnav, E. D. et al. The evolution, evolvability and engineering of gene regulatory DNA. Nature 603, 455–463 (2022).
Article CAS PubMed ADS Google Scholar
Venturelli, O. S. et al. Deciphering microbial interactions in synthetic human gut microbiome communities. Mol. Syst. Biol. 14, e8157 (2018).
Article PubMed PubMed Central Google Scholar
Michel-Mata, S., Wang, X. W., Liu, Y. Y. & Angulo, M. T. Predicting microbiome compositions from species assemblages through deep learning. iMeta 1, e3 (2022).
Article PubMed PubMed Central Google Scholar
Wang, X. W. et al. Identifying keystone species in microbial communities using deep learning. Nat. Ecol. Evol. 8, 22–31 (2023).
Article PubMed Google Scholar
Bashan, A. et al. Universality of human microbial dynamics. Nature 534, 259–262 (2016).
Article CAS PubMed PubMed Central ADS Google Scholar
Bunin, G. Ecological communities with Lotka-Volterra dynamics. Phys. Rev. E 95, 042414 (2017).
Article MathSciNet PubMed ADS Google Scholar
Hernandez-Sanabria, E., Vazquez-Castellanos, J. F. & Raes, J. In vitro ecology: a discovery engine for microbiome therapies. Nat. Rev. Gastroenterol. Hepatol. 17, 711–712 (2020).
Article PubMed Google Scholar
Javdan, B. et al. Personalized mapping of drug metabolism by the human gut microbiome. Cell 181, 1661–1679.e22（2020).
Li, L. et al. RapidAIM: a culture- and metaproteomics-based Rapid Assay of Individual Microbiome responses to drugs. Microbiome 8, 33 (2020).
Article CAS PubMed PubMed Central Google Scholar
Aranda-Diaz, A. et al. Establishment and characterization of stable, diverse, fecal-derived in vitro microbial communities that model the intestinal microbiota. Cell Host Microbe 30, 260–272 e265 (2022).
Article CAS PubMed PubMed Central Google Scholar
Hanchi, H., Mottawea, W., Sebei, K. & Hammami, R. The genus Enterococcus: between probiotic potential and safety concerns—an update. Front Microbiol. 9, 1791 (2018).
Article PubMed PubMed Central Google Scholar
Griffin, M. E. et al. Enterococcus peptidoglycan remodeling promotes checkpoint inhibitor cancer immunotherapy. Science 373, 1040–1046 (2021).
Article CAS PubMed PubMed Central ADS Google Scholar
Xiong, X. et al. Emerging enterococcus pore-forming toxins with MHC/HLA-I as receptors. Cell 185, 1157–1171 e1122 (2022).
Article CAS PubMed PubMed Central Google Scholar
Van Tyne, D. & Gilmore, M. S. Friend turned foe: evolution of enterococcal virulence and antibiotic resistance. Annu. Rev. Microbiol. 68, 337–356 (2014).
Article PubMed PubMed Central Google Scholar
Jones, M. L., Rivett, D. W., Pascual-Garcia, A. & Bell, T. Relationships between community composition, productivity and invasion resistance in semi-natural bacterial microcosms. eLife 10, e71811 (2021).
Article CAS PubMed PubMed Central Google Scholar
Hromada, S. et al. Negative interactions determine Clostridioides difficile growth in synthetic human gut communities. Mol. Syst. Biol. 17, e10355 (2021).
Article CAS PubMed PubMed Central Google Scholar
Keesing, F. et al. Impacts of biodiversity on the emergence and transmission of infectious diseases. Nature 468, 647–652 (2010).
Article CAS PubMed PubMed Central ADS Google Scholar
Valles-Colomer, M. et al. Variation and transmission of the human gut microbiota across multiple familial generations. Nat. Microbiol. 7, 87–96 (2022).
Article CAS PubMed Google Scholar
Derosa, L. et al. Intestinal Akkermansia muciniphila predicts clinical response to PD-1 blockade in patients with advanced non-small-cell lung cancer. Nat. Med. 28, 315–324 (2022).
Article CAS PubMed PubMed Central Google Scholar
Dao, M. C. et al. Akkermansia muciniphila and improved metabolic health during a dietary intervention in obesity: relationship with gut microbiome richness and ecology. Gut 65, 426–436 (2016).
Article CAS PubMed Google Scholar
Zhai, Q. X., Feng, S. S., Arjan, N. & Chen, W. A next generation probiotic, Akkermansia muciniphila. Crit. Rev. Food Sci. 59, 3227–3236 (2019).
Article CAS Google Scholar
Karcher, N. et al. Genomic diversity and ecology of human-associated Akkermansia species in the gut microbiome revealed by extensive metagenomic assembly. Genome Biol. 22, 209 (2021).
Article CAS PubMed PubMed Central Google Scholar
Ribeiro, M. T., Singh, S. & Guestrin, C. in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 1135–1144 (2016).
Chen, J. et al. A commensal-encoded genotoxin drives restriction of Vibrio cholerae colonization and host gut microbiome remodeling. Proc. Natl Acad. Sci. USA 119, e2121180119 (2022).
Article CAS PubMed PubMed Central Google Scholar
Weiss, A. S. et al. In vitro interaction network of a synthetic gut bacterial community. ISME J. 16, 1095–1109 (2022).
Article CAS PubMed Google Scholar
Getzke, F. et al. Cofunctioning of bacterial exometabolites drives root microbiota establishment. Proc. Natl Acad. Sci. USA 120, e2221508120 (2023).
Article CAS PubMed PubMed Central Google Scholar
Brugiroux, S. et al. Genome-guided design of a defined mouse microbiota that confers colonization resistance against Salmonella enterica serovar Typhimurium. Nat. Microbiol. 2, 16215 (2016).
Article CAS PubMed Google Scholar
Kehe, J. et al. Massively parallel screening of synthetic microbial communities. Proc. Natl Acad. Sci. USA 116, 12804–12809 (2019).
Article CAS PubMed PubMed Central ADS Google Scholar
Gopalakrishnappa, C., Gowda, K., Prabhakara, K. H. & Kuehn, S. An ensemble approach to the structure-function problem in microbial communities. iScience 25, 103761 (2022).
Article CAS PubMed PubMed Central ADS Google Scholar
Xiao, Y., Angulo, M. T., Lao, S., Weiss, S. T. & Liu, Y. Y. An ecological framework to understand the efficacy of fecal microbiota transplantation. Nat. Commun. 11, 3329 (2020).
Article CAS PubMed PubMed Central ADS Google Scholar
Mayfield, M. M. & Stouffer, D. B. Higher-order interactions capture unexplained complexity in diverse communities. Nat. Ecol. Evol. 1, 62 (2017).
Article PubMed Google Scholar
Sanchez-Gorostiaga, A., Bajic, D., Osborne, M. L., Poyatos, J. F. & Sanchez, A. High-order interactions distort the functional landscape of microbial consortia. PLoS Biol. 17, e3000550 (2019).
Article CAS PubMed PubMed Central Google Scholar
Debray, R. et al. Priority effects in microbiome assembly. Nat. Rev. Microbiol. 20, 109–121 (2022).
Article CAS PubMed Google Scholar
Rao, C. et al. Multi-kingdom ecological drivers of microbiota assembly in preterm infants. Nature 591, 633–638 (2021).
Article CAS PubMed PubMed Central ADS Google Scholar
Gibson, T. E., Bashan, A., Cao, H.-T., Weiss, S. T. & Liu, Y.-Y. On the origins and control of community types in the human microbiome. PLoS Comput. Biol. 12, e1004688 (2016).
Article PubMed PubMed Central Google Scholar
You, H. J. et al. Bacteroides vulgatus SNUG 40005 restores Akkermansia depletion by metabolite modulation. Gastroenterology 164, 103–116 (2023).
Article CAS PubMed Google Scholar
Sanchez, A. et al. The community-function landscape of microbial consortia. Cell Syst. 14, 122–134 (2023).
Article CAS PubMed Google Scholar
Skwara, A. et al. Statistically learning the functional landscape of microbial communities. Nat. Ecol. Evol. 7, 1823–1833 (2023).
Article PubMed Google Scholar
Vrancken, G., Gregory, A. C., Huys, G. R. B., Faust, K. & Raes, J. Synthetic ecology of the human gut microbiota. Nat. Rev. Microbiol 17, 754–763 (2019).
Article CAS PubMed Google Scholar
Vogel, C. M., Potthoff, D. B., Schafer, M., Barandun, N. & Vorholt, J. A. Protective role of the Arabidopsis leaf microbiota against a bacterial pathogen. Nat. Microbiol. 6, 1537–1548 (2021).
Article CAS PubMed PubMed Central Google Scholar
Herrera Paredes, S. et al. Design of synthetic bacterial communities for predictable plant phenotypes. PLoS Biol. 16, e2003962 (2018).
Article PubMed PubMed Central Google Scholar
van de Velde, C. et al. Technical versus biological variability in a synthetic human gut community. Gut microbes 15, 2155019 (2023).
Article PubMed Google Scholar
Hu, H. et al. StrainPanDA: Linked reconstruction of strain composition and gene content profiles via pangenome‐based decomposition of metagenomic data. iMeta 1 e41 (2022).
Osbelt, L. et al. Klebsiella oxytoca causes colonization resistance against multidrug-resistant K. pneumoniae in the gut via cooperative carbohydrate competition. Cell Host Microbe 29, 1663–1679 e1667 (2021).
Article CAS PubMed Google Scholar
Wu, S. et al. GMrepo: a database of curated and consistently annotated human gut metagenomes. Nucleic Acids Res. 48, D545–D553 (2020).
Article CAS PubMed Google Scholar
Gonze, D., Lahti, L., Raes, J. & Faust, K. Multi-stability and the origin of microbial community types. ISME J. 11, 2159–2166 (2017).
Article PubMed PubMed Central Google Scholar
Dai, L., Vorselen, D., Korolev, K. S. & Gore, J. Generic indicators for loss of resilience before a tipping point leading to population collapse. Science 336, 1175–1177 (2012).
Article CAS PubMed ADS Google Scholar
Davey, L. E. et al. A genetic system for Akkermansia muciniphila reveals a role for mucin foraging in gut colonization and host sterol biosynthesis gene expression. Nat. Microbiol. 8, 1450–1467 (2023).
Schubert, A. M., Sinani, H. & Schloss, P. D. Antibiotic-induced alterations of the murine gut microbiota and subsequent effects on colonization resistance against Clostridium difficile. mBio 6, e00974 (2015).
Article CAS PubMed PubMed Central Google Scholar
Clark, R. L. et al. Design of synthetic human gut microbiome assembly and butyrate production. Nat. Commun. 12, 3254 (2021).
Article CAS PubMed PubMed Central ADS Google Scholar
Baxter, N. T. et al. Dynamics of human gut microbiota and short-chain fatty acids in response to dietary interventions with three fermentable fibers. mBio 10, e02566–18 (2019).
Article CAS PubMed PubMed Central Google Scholar
Suez, J. et al. Personalized microbiome-driven effects of non-nutritive sweeteners on human glucose tolerance. Cell 185, 3307–3328 e3319 (2022).
Article CAS PubMed Google Scholar
Li, L. et al. An in vitro model maintaining taxon-specific functional activities of the gut microbiome. Nat. Commun. 10, 4146 (2019).
Article CAS PubMed PubMed Central ADS Google Scholar
Browne, H. P. et al. Culturing of ‘unculturable’ human microbiota reveals novel taxa and extensive sporulation. Nature 533, 543 (2016).
Article CAS PubMed PubMed Central ADS Google Scholar
Hillmann, B. et al. Evaluating the information content of shallow shotgun metagenomics. mSystems 3, e00069–18 (2018).
Article CAS PubMed PubMed Central Google Scholar
Maier, L. et al. Unravelling the collateral damage of antibiotics on gut bacteria. Nature 599, 120–124 (2021).
Article CAS PubMed PubMed Central ADS Google Scholar
EUCAST, T. European Committee on Antimicrobial Susceptibility Testing, Breakpoint tables for interpretation of MICs and zone diameters. Version 5.0, 2015 (2015).
Huang, S. et al. Candidate probiotic Lactiplantibacillus plantarum HNU082 rapidly and convergently evolves within human, mice, and zebrafish gut but differentially influences the resident microbiome. Microbiome 9, 151 (2021).
Article CAS PubMed PubMed Central Google Scholar
Zhang, Z. et al. Lactobacillus fermentum HNU312 alleviated oxidative damage and behavioural abnormalities during brain development in early life induced by chronic lead exposure. Ecotoxicol. Environ. Saf. 251, 114543 (2023).
Article CAS PubMed Google Scholar
Cani, P. D. & de Vos, W. M. Next-generation beneficial microbes: the case of Akkermansia muciniphila. Front. Microbiol. 8, 1765 (2017).
Article PubMed PubMed Central Google Scholar
Xu, W. et al. Characterization of shallow whole-metagenome shotgun sequencing as a high-accuracy and low-cost method by complicated mock microbiomes. Front. Microbiol. 12, 678319 (2021).
Article PubMed PubMed Central Google Scholar
Clarke, K. R. Non‐parametric multivariate analyses of changes in community structure. Aust. J. Ecol. 18, 26 (1993).
Article Google Scholar
Dixon, P. VEGAN, a package of R functions for community ecology. J. Veg. Sci. 14.6, 3 (2003).
Google Scholar
Segata, N. et al. Metagenomic biomarker discovery and explanation. Genome Biol. 12, R60 (2011).
Article PubMed PubMed Central Google Scholar

Download references

Acknowledgements

We thank Na Li and volunteers at Shenzhen Institute of Advanced Technology (SIAT) for stool sample collection. We thank Zepeng Qu and Zhenkun Zhang for isolating human gut bacterial strains. We thank Prof. Jiachao Zhang at Hainan University for providing probiotic strains. We thank Huaijie Hao and Yan Tan at Xbiome for providing help with the anaerobic workstation. We thank Shenzhen Infrastructure for Synthetic Biology and Chaobi Lei for providing support in DNA extraction. We thank Zheng Sun, Chen Liao, Hongbin Liu, Lanxiang Wang, and colleagues at SIAT for valuable discussions. L.D. acknowledges support from the National Key R&D Program of China (2019YFA0906700), the National Natural Science Foundation of China (31971513), and Shenzhen Key Laboratory for the Intelligent Microbial Manufacturing of Medicines (ZDSYS20210623091810032). Y.-Y.L. is supported by grants R01AI141529, R01HD093761, RF1AG067744, UH3OD023268, U19AI095219 and U01HL089856 from the National Institutes of Health, USA; a pilot grant from the Biology of Trauma Initiative of Broad Institute, USA; and the Office of the Assistant Secretary of Defense for Health Affairs, through the Traumatic Brain Injury and Psychological Health Research Program (Focused Program Award) under award no. (W81XWH-22-S-TBIPH2), endorsed by the Department of Defense, USA. L.W. acknowledges support from the National Natural Science Foundation of China (32100089). X.W.W. acknowledges the funding support from the National Institutes of Health (K25HL166208).

Author information

These authors contributed equally: Lu Wu, Xu-Wen Wang.

Authors and Affiliations

CAS Key Laboratory of Quantitative Engineering Biology, Shenzhen Institute of Synthetic Biology, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China
Lu Wu, Zining Tao, Wenlong Zuo, Yu Zeng & Lei Dai
Channing Division of Network Medicine, Department of Medicine, Brigham and Women’s Hospital and Harvard Medical School, Boston, MA, USA
Xu-Wen Wang, Tong Wang & Yang-Yu Liu
Shandong Agricultural University, Tai’an, China
Zining Tao
Center for Artificial Intelligence and Modeling, The Carl R. Woese Institute for Genomic Biology, University of Illinois at Urbana-Champaign, Champaign, IL, USA
Yang-Yu Liu
University of Chinese Academy of Sciences, Beijing, China
Lei Dai

Authors

Lu Wu
View author publications
You can also search for this author in PubMed Google Scholar
Xu-Wen Wang
View author publications
You can also search for this author in PubMed Google Scholar
Zining Tao
View author publications
You can also search for this author in PubMed Google Scholar
Tong Wang
View author publications
You can also search for this author in PubMed Google Scholar
Wenlong Zuo
View author publications
You can also search for this author in PubMed Google Scholar
Yu Zeng
View author publications
You can also search for this author in PubMed Google Scholar
Yang-Yu Liu
View author publications
You can also search for this author in PubMed Google Scholar
Lei Dai
View author publications
You can also search for this author in PubMed Google Scholar

Contributions

L.D., Y.Y.L., L.W., and X.W.W. conceived the presented idea, designed the project, interpreted the results, and wrote the manuscript. X.W.W. implemented in silico simulations, machine learning models, and analytical derivations. L.W. and Z.T. performed the in vitro experiments. Y.Z. assisted in experiments. L.W. analyzed the experimental data. T.W. discussed the results and revised the manuscript. W.Z. assisted in data analysis and interpretation. All authors approved the final manuscript. L.D. and Y.Y.L. provided supervision and resources for this study.

Corresponding authors

Correspondence to Yang-Yu Liu or Lei Dai.

Ethics declarations

Competing interests

The authors declare no competing interests.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Supplementary Information

Reporting Summary

Supplementary Data 1

Supplementary Data 2

Description of Additional Supplementary Files

Peer Review File

Source data

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Cite this article

Wu, L., Wang, XW., Tao, Z. et al. Data-driven prediction of colonization outcomes for complex microbial communities. Nat Commun 15, 2406 (2024). https://doi.org/10.1038/s41467-024-46766-y

Download citation

Received: 07 July 2023
Accepted: 08 March 2024
Published: 16 March 2024
DOI: https://doi.org/10.1038/s41467-024-46766-y

Comments

By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.