To determine the nature of time series statistics generated by resource competition, we considered a minimal CR model in which N consumers compete for M resources via growth dynamics described bydXidt=Xi∑j=1MRijYj,(1)dYjdt=−Yj∑i=1NRijXi.

Here, Xi denotes the abundance of consumer i, Yj the amount of resource j, and Rij the consumption rate of resource j by consumer i. The resources in this model are defined at a coarse-grained level, such that individual resources represent effective groups of metabolites or niches. We assumed that the resource consumption rates Rij were independent of the external environment and constant over time, thereby specifying the intrinsic ecological properties of the community with a collection of N×M microscopic parameters. To simplify this vast parameter space, we conjectured that the macroecological features of our experimental time series might be captured by typical profiles of resource consumption drawn from a statistical ensemble. This is a crucial simplification: while these randomly drawn values will never match the specific resource consumption rates of a given microbiota, previous work suggests that they can often recapitulate the large-scale behavior of sufficiently diverse communities (Cui et al., 2021). This simplification allows us to test whether particular ensembles of resource consumption rates can reproduce the time series statistics we observe. Specifically, we considered an ensemble in which each Rij was randomly selected from a uniform distribution between 0 and Rmax. To model the sparsity of resource competition within the community, each Rij was set to zero with probability S (Figure 1A). This ensemble approach allows us to represent arbitrarily large communities with just two global parameters, S and Rmax.

We simulated the dynamics in Equation 1 using a serial dilution scheme (Erez et al., 2020) to mimic the punctuated turnover of gut microbiotas due to multiple feedings and defecations between sampling times. During a sampling interval T, each dilution cycle was seeded with an initial amount of each resource, Yj,0(T), and Equation 1 was simulated until all resources were depleted (dYj/dt=0 for all j). The community was then diluted by a factor D and resources were replenished to their initial amounts Yj,0T (Figure 1B). To mimic the effects of a reservoir of species that could potentially compete for the resources (Ng et al., 2019), we initialized the first dilution cycle of each sampling interval by assuming that N consumers were present at equal abundance. Additional dilution cycles were then performed until an approximate ecological steady state was reached (Figure 1B, Materials and methods). Consumer abundances at sampling time T were defined by this approximate ecological steady state. For the relevant parameter regimes we considered, this approximate steady state was reached within a reasonable number of generations (5–6 dilutions or ~40 generations for D=200). Although the precise details of microbiota turnover are largely unknown in humans, our modeling results were robust to the precise value of D and threshold for ecological steady state (Figure 1—figure supplement 1). Similarly, our results did not depend on the precise composition of the reservoir (Figure 1—figure supplement 2), although they did depend on its existence and relative size (Figure 1—figure supplement 2).

Under the assumptions of this model, any temporal variation in consumer abundances must arise through external fluctuations in the initial resource levels Yj,0T, which might come, for example, from dietary fluctuations. To model these fluctuations, we assumed that the initial resource levels undergo a biased random walk around their average values Yj-:(2)Yj,0(T)=|Yj,0(T−1)−k(Yj,0(T−1)−Yj−)+σYj−ξj(T)|,

where ξjT is a normally distributed random variable with zero mean and unit variance, σ determines the magnitude of resource fluctuations, and k is the strength of a restoring force that ensures the same resource environment on average over time (Figure 1A). The absolute value enforces Yj,0 to be positive. If k=0, there is no restoring force and hence Yj,0T performs an unbiased random walk; if k=1, Yj,0T fluctuates about its set point Yj¯ independent of its value at the previous sampling time. For all k>0, the model exhibits long-term stability without drift. As above, we used an ensemble approach to model the set points Yj¯, assuming that each Yj¯ was independently drawn from a uniform distribution between 0 and Ymax. These assumptions yield a Markov chain of fluctuating resource amounts Yj,0T and their corresponding consumer relative abundances xi(T)=Xi(T)/∑nXn(T) (Figure 1C).

The statistical properties of these time series are primarily determined by five global parameters: the total number of consumers in the reservoir N, the number of resources in the environment M, the sparsity S of the resource consumption matrix, and the resource fluctuation parameters σ and k. The absolute magnitudes of Rmax and Ymax are not important for our purposes since they do not affect the predictions of consumer relative abundances at ecological steady state. We extracted N from experimental data as the number of consumers that were present for at least one sampling time point, leaving only four free global parameters.

Previous studies have suggested that the family level is an appropriate coarse graining of metabolic capabilities (Goldford et al., 2018; Louca et al., 2016; Tian et al., 2020), thus we assumed, unless otherwise specified, that each consumer grouping i within our model represents a taxonomic family, and combined abundances of empirical operational taxonomic units (OTUs) or amplicon sequencing variants (ASVs; Callahan et al., 2016) at the family level for analyses (Materials and methods). Given the typical limits of detection of 16S amplicon sequencing data sets, we only examined time series statistics for taxa with relative abundance >10^–4 at any given time point. Experimental and simulated data were processed equivalently to enable consistent comparisons of their time series statistics.

As expected, we found that random realizations of our model (i.e., different resource consumption matrices drawn from the same ensemble) generated similar time series statistics, whose typical behavior strongly varied with the global parameters of the model. In particular, only small subsets of the parameters led to time series statistics that agreed with experiments, as we show below. An example simulation using the macroscopic parameters N,M,S,σ, k=50, 30, 0.1, 0.2, 0.8 is shown in Figure 1. This set of parameters produced relative abundance time series with highly similar statistical behaviors as in experiments involving daily sampling of human stool (Figure 1D). Given this agreement, we next systematically analyzed the time series statistics generated by our model across the macroscopic parameter space and compared against experimental behaviors to estimate model parameters for wide-ranging microbiotas.

To test whether our model can recapitulate major features of experimental time series, we first focused on a data set of daily sampling of the gut microbiota from a human subject (Caporaso et al., 2011; Figure 2). These data were previously shown (Ji et al., 2020) to exhibit several distinctive statistical behaviors: (1) the variance σxi2 of family i over the sampling period scaled as a power law with its mean xi (Figure 2B and F); (2) the log₁₀(abundance change) ΔliT=log10⁡xiT+1/xiT , pooled over all families and across all sampling times, was well fit by an exponential distribution with standard deviation σΔl (Figure 2B and G); and (3) the distributions of residence times tres and return times tret (the durations of sustained presence and absence, respectively) pooled over all families were well fit by power laws with an exponential cutoff (Figure 2D and K). Through an exhaustive search of parameter space, we identified a specific combination of parameters that could reproduce all of these behaviors within our simple CR model (Figure 2F, G and K).

In addition, several other important statistics were reproduced without any additional fitting: (1) the distribution of richness αT , the number of consumers present at sampling time T (Figure 2A and E); (2) the distribution of the restoring slopes si of the linear regression of ΔliT against liT≡log10⁡xiT across all T (Figure 2C and H); (3) the distribution of prevalences pi , the fraction of sampling times for which family i is present (Figure 2A1); (4) the relationship between pi and xi (Figure 2J); and (5) the rank distribution of mean abundances xi (Figure 2L).

Therefore, our model was able to simultaneously capture at least eight statistical behaviors in a microbiota time series with only four parameters, each of which may represent biologically relevant features of the community.

To determine whether our model can be used to analyze time series statistics at other taxonomic levels, we analyzed the same data set (Caporaso et al., 2011) at finer (genus) and coarser (class) taxonomic levels, both of which exhibited qualitatively similar statistical behaviors as the family level. Our modeling framework was able to quantitatively recapitulate almost all statistics at both levels (Figure 2—figure supplements 1 and 2). A notable exception is that the Bacteroides genus dominated the observed rank abundance distribution at the genus level, while our CR model predicted a more even distribution (Figure 2—figure supplement 2). Nevertheless, the relative abundances among the remaining genera were still well captured by the model predictions (Figure 2—figure supplement 2). These results demonstrate that our model and its applications can be generalized across taxonomic levels.

Since our model can reproduce the observed statistics in gut microbiota time series, we sought to determine how these statistics would respond to changes in model parameters, and thus how experimental measurements constrain the ensemble parameters across various data sets. To do so, we simulated our model across all relevant regions of parameter space. S and k were varied across their entire ranges, and M and σ were varied across relevant regions outside of which the model clearly disagreed with the observed data. For each set of parameters, each time series statistic was averaged across random instances of Rij and Yj,0T drawn from the same statistical ensemble. For each statistic z, its global susceptibility C(z,w) to parameter w was calculated as the change in z when w is varied, averaged over all other parameters and normalized by the standard deviation of z across the entire parameter space. Due to the normalization, Cz,w varies approximately between –3 and 3, where a magnitude close to 3 indicates that almost all the variance of z is due to changing w.

By clustering and ranking susceptibilities, we identified four statistics with |C(z,w)|>2 that were largely determined by one of each of the four model parameters (Figure 3, Figure 3—figure supplement 1): mean richness α, the power law exponent β of σxi2 versus xi, the standard deviation in log₁₀(abundance change) σΔl, and the mean restoring slope s were almost exclusively susceptible to variations in M, S, σ, and k, respectively. Similar results were also obtained for local versions of the susceptibility, in which individual parameters were varied around the best fit values for the human gut microbiota in Figure 2 (Figure 3-figure supplement 2). These susceptibilities broadly illustrate how various time series statistics are affected by coarse-grained parameters of resource competition; we further investigate some specific examples in the next section.

The exclusive susceptibilities of these four statistics suggest that they can serve as informative metrics for estimating model parameters. Therefore, we estimated model parameters by minimizing the sum of errors between model predictions and experimental measurements of these four statistics, and obtained estimation bounds by determining parameter variations that would increase model error by 5% of the mean error across all parameter space. As we will show, the resulting bounds are small relative to the differences among distinct microbiotas, indicating that meaningful conclusions can be drawn from the best fit values of the ensemble level parameters of resource competition. In summary, the four model parameters were fit to four summary statistics: mean richness α, variance-mean scaling exponent β, standard deviation of abundance change σΔl, and mean restoring slope s (Figure 2E–H, respectively). The shapes of their corresponding distributions and scalings, as well as at least four other statistics (Figure 2I–L), are all parameter-free predictions.

To understand the mechanisms that underlie the susceptibilities of various time series statistics to model parameters, we investigated their origins within our model, focusing on how they constrain the parameters.

The average richness α is a fundamental descriptor of community diversity. Within our model, α is largely determined by and increases with increasing resource number M (Cα,N/M=-2.6), as expected for CR dynamics. The sparsity of resource use S impacts the power law exponent β between σxi2 and xi (Cβ,S=-2.0). Together, α and β constrain the parameters of resource competition M and S.

The effect of S on β can be partially understood by considering limiting behaviors as follows. When sparsity is high (S≈1), there is little competition and each consumer consumes almost distinct sets of resources from other consumers. In the limit in which each consumer utilizes a single unique resource, σxi2 is determined by the noise in resource level, which has a β=2 scaling according to Equation 2. In the limit of large M and high sparsity, the variation in the number of resources consumed by each consumer can be large relative to the mean, and both σxi2 and xi scale with the number of resources consumed, hence β=1. Simulations of a no-competition model in which consumers consume distinct sets of resources confirmed the scalings in these limits (Figure 3—figure supplement 3). By contrast, when sparsity is low (S≈0), each consumer utilizes almost all resources and hence variation in the number of resources consumed is small relative to the mean. Despite the obvious presence of competition in our CR model, we nevertheless attempted to understand the low sparsity limit by extrapolating the no-competition model above to a case in which all consumers consume distinct sets of the same number of resources. For large number of resources, these simulations predicted that β≈1.5 (Figure 3—figure supplement 3), as did our CR model for S=0.1 (Figure 3—figure supplement 3). These findings suggest that the effect of S on β can be partially attributed to differences in the number of resources consumed.

The distribution of Δl describes the nature of abundance changes. As expected, the width of the distribution is largely determined by and increases with increasing σ (CσΔl,σ=2.6). For the gut microbiota data set in Figure 2, the shape of the distribution was well fit by an exponential. Within our model, the shape of the distribution aggregated across all consumers is determined by N/M and the sparsity S, emerging from the mixture of each consumer’s individual distribution (Figure 3—figure supplement 4). When N/M<1 and the sparsity S is low, individual distributions of Δl are well fit by normal distributions, and pool together to generate another normal distribution. When N/M<1 and sparsity S is high, individual distributions remain normal, but can pool together to generate a non-normal distribution that is well fit by an exponential (see also Allen et al., 2001). By contrast, when N/M>1, individual distributions can be well fit by an exponential and can pool together to approximate another exponential. Simulations of the no-competition model considered above led to individual and aggregate distributions that were normal in all cases, indicating that in our model resource competition is responsible for generating the non-normal distributions of Δl (Figure 3—figure supplement 3). Although it is challenging to discern the shape of individual distributions in most experimental data sets given the limited numbers of samples, the shape of the aggregate distribution of Δl informs the parameters of resource competition M and S. In particular, an exponential distribution of Δl suggests either strong resource competition in the form of N>M or substantial niche differentiation in the form of high S. Other statistics such as β can help to distinguish between these two regimes.

The distribution of restoring slopes si describes the tendency with which consumers revert to their mean abundances following fluctuations. As expected, the mean s is almost completely determined by k, which describes the autocorrelation in resource levels (-s≈k and Cs,k=-3.0). Together, the distributions of Δl and si constrain the parameters of external fluctuations σ and k.

Within our model, resource fluctuations can lead to the temporary ‘extinction’ of certain species when they drop below the detectability threshold of 10^–4. The distributions of residence and return times, tres and tret, reflect the probabilities of extinction as well as correlations between sampling times. For all parameter sets explored, these distributions can be well fit by power laws, with an exponential cutoff to account for finite sampling (Ji et al., 2020). As expected, the power law slopes νres and νret decrease (become more negative) with increasing σ or k (Figure 3—figure supplement 1), since increasing external noise or decreasing correlations in time increases the probability of fluctuating between existence and extinction for each consumer. By contrast, νres and νret change in opposite directions in response to variation in M (Figure 3—figure supplement 1). Increasing M leads to a larger number of highly prevalent consumers, thereby increasing the mean and broadening the distribution of tres and decreasing the mean and narrowing the distribution of tret. Since the four ensemble level parameters are already fixed by other statistics, the distributions of tres, tret, and pi are parameter-free predictions of our model. In other words, a macroscopic characterization of the effective resource competition and resource fluctuations is sufficient to predict the statistics of ‘extinction’ dynamics, as well as the abundance rank distribution and the relationship between consumer abundance and prevalence.

Since the distributions of Δl, tres, and tret are dependent on correlations between sampling times, it was initially puzzling that their distributions in some data sets remained similar after shuffling sampling times, raising questions as to what extent these statistics hold information about the underlying intrinsic dynamics (Tchourine et al., 2021; Wang and Liu, 2021a). Our results assist in reconciling the apparent conundrum, since within our model richness α and Taylor’s law exponent β do not depend on correlations between sampling times and are also the statistics that are most informative about the intrinsic parameters M and S (Figure 3). As a result, the shuffled time series were also well fit by our model and yielded best fit values that were identical to those produced by the actual time series except with k=1, as expected due to the absence of correlation across sampling times (Figure 3—figure supplement 5). Thus, our results suggest that while external fluctuations in resource levels may be responsible for generating species abundance variations, the intrinsic properties of resource competition can determine the resulting scaling exponents of many statistical behaviors.

Taken together, our analyses demonstrate the complex relationships among time series statistics and highlight their unification within our model using only a small number of global parameters, whose values are strongly constrained by macroecological patterns.

We have shown that many time series statistics can be recapitulated by a simple model that does not require knowing many detailed features of real microbiota (Figure 2). The success of this approach implies that these macroecological fluctuations must be independent of at least some model details, which suggests that there may be other ecological models that could also recapitulate the same data (Figure 4, Figure 4—figure supplements 1–5). The relationships between ecological models are generally poorly characterized. To explore these possibilities, we sought to compare our calibrated CR models against several common alternatives.

First, we aimed to determine the extent to which the simulated statistics depend on the assumptions of our CR model. Our parametrization of the consumption rates introduces a correlation between the maximum growth rate of a consumer and the number of resources it consumes. To remove this correlation, we normalized the sum of consumption rates ∑jRij for consumer i to a fixed capacity R~i that was randomly drawn from the original growth rates ∑jRijYj,0 (Good et al., 2018; Posfai et al., 2017; Tikhonov and Monasson, 2017). This modification preserves the variation in consumer fitness while implementing a metabolic trade-off. The resulting time series statistics were essentially unaffected, also recapitulating experimental data (Figure 4—figure supplement 1).

Moreover, the CR dynamics in Equation 1 do not consider other biologically plausible scenarios such as saturation kinetics (Momeni et al., 2017; Niehaus et al., 2019). To probe the robustness of the results of our model to the dynamical assumptions, we implemented saturation kinetics with all other details kept the same (Materials and methods). When this model was simulated with the best fit parameters of the original model, the resulting dynamics were less variable across sampling times than without saturation kinetics, since the saturated regime is unaffected by small changes in resource levels (Figure 4—figure supplement 2). Nonetheless, experimental statistics were again reproduced once the strength of environmental fluctuations σ was increased appropriately (Figure 4—figure supplement 2). This suggests that our results are robust to assumptions regarding metabolic trade-offs and saturation kinetics.

We next considered a non-interacting null model in which consumer abundances were drawn from independent normal distributions whose means and variances were fitted directly from the data. Even with a large number of free parameters, this null model was unable to capture some of the time series statistics reproduced by our CR model, including Taylor’s law as well as the distributions of richness and restoring slopes (Figure 4—figure supplement 3). We reasoned that the discrepancies between experimental data and the null model could be due to the lack of interspecies interactions. To test this hypothesis, we examined the pairwise correlations between the consumer abundances across sampling times. The measured distribution of pairwise correlations is much broader than the prediction of the non-interacting model, which is sharply peaked about zero as expected (Figure 4). By contrast, the distribution of correlations predicted by our CR model without any additional fitting was in much closer agreement with the experimental data (Figure 4). These findings imply that interspecies interactions are required to capture important details of community dynamics.

While our CR model assumes pairwise interactions between consumers and resources, the effective interactions between consumers are not necessarily pairwise. To explore whether these higher-order contributions are necessary for recapitulating the data, we considered models explicitly based on pairwise interspecies interactions, which despite differences compared with CR models (Momeni et al., 2017) can also reproduce some properties of experimental time series (Descheemaeker and de Buyl, 2020; Wang and Liu, 2021b). To further explore the properties of models focused on pairwise interactions, we investigated gLV models in which N taxa grow and interact via(3)dXidt=Xi(ri+∑j=1NAijXj−Γ(t)),

where Xi denotes the relative abundance of taxon i, ri its growth rate, and Aij its interaction coefficient with taxon j. Γ(t)=∑iriXi+∑i,jAijXiXj is a normalizing term that ensures that the relative abundances always sum to one (Joseph et al., 2020). Since this classical model is generally unstable for randomly drawn interaction coefficients (May, 1972), we sought to focus on particular instances of the gLV model that were closest to our original CR model. This conversion between models was achieved by converting the consumption rates Rij and resource levels Yj,0 at each sampling time T to the growth rates ri and interaction coefficients Aij that characterize the dynamics when consumption rates are similar to the mean value (Materials and methods). This conversion results in negative, symmetric Aij whose magnitudes depend on the niche overlap between the interacting taxa (Good et al., 2018). Moreover, fluctuations in Yj,0 result in corresponding fluctuations in both ri and Aij across T. These CR-converted gLV models generated time series statistics that reproduced the experimental data to a similar extent as the original CR model (Figure 4—figure supplement 4). In light of this correspondence, we asked whether more general ensembles of pairwise interaction could also reproduce the experimental data. We randomly selected ri and Aij values from normal distributions with means and variances equal to those in the CR-converted gLV models while enforcing symmetric and negative interactions. The resulting gLV models yielded a poor fit to the data (Figure 4—figure supplement 5). Together, these results suggest that while pairwise interactions between taxa are likely sufficient to recapitulate the experimental data, their parameters must be drawn from particular ensembles that can be more simply described in the CR framework.

These examples reinforce that only a particular subset of models can recapitulate the data, and therefore, that the underlying community properties are highly constrained by macroecological dynamics. Moreover, our calibrated CR model can guide the parametrization of other models that can satisfy those constraints, while also identifying model features that are necessary for recapitulating data.

Having developed a simple method to estimate parameters of our CR model that recapitulate time series statistics, we applied this method to data sets involving wide-ranging microbial communities. Although the various communities considered are drastically different in many aspects, we hypothesized that our CR model framework could still be applied to identify the statistical ensembles that can describe their macroecological dynamics. In addition to microbiotas from the human and mouse gut (Caporaso et al., 2011; Carmody et al., 2015; David et al., 2014), we examined communities from the human vagina (Song et al., 2020), human saliva (David et al., 2014), and in and around rice roots (Edwards et al., 2018). The time series statistics of these microbiotas varied broadly (Figure 5A). Nevertheless, our model successfully reproduced the experimental statistics across all communities (Figure 5—figure supplements 1–6), suggesting that simple CR models can capture many of the macroscopic features of these microbiotas.

The best fit parameters suggest that the effective resource competition dynamics occur in distinct regimes across microbiotas (Figure 5B). Human gut microbiotas were best described by N>M, suggesting that there are more species in the reservoir than resources in the environment, by contrast to mouse gut microbiotas that were best described by N<M. In terms of resource niche overlaps, human gut microbiotas were best fit with sparsity S<0.3, while mouse gut microbiotas were best fit with S>0.3, suggesting that on average, pairs of bacterial families are more metabolically distinct in the mouse versus the human gut.

Unlike gut microbiotas, a human saliva microbiota yielded best fit parameters N≈M and S≈0.8, suggesting that this community has access to abundant resources and that each effective resource is competed for by a small fraction of the extant bacterial families. All vaginal microbiotas were best fit with S<0.1, suggesting intense resource competition.

Like vaginal microbiotas, microbial communities residing in the bulk soil around rice roots and in the associated rhizoplane and rhizosphere were well described by S<0.1. By contrast, the community in the associated endosphere was best described by S≈0.6, suggesting that resource competition is less fierce within plant roots than around them.

In addition, inferences about the nature of environmental fluctuations can be made from the best fit values of σ and k (Figure 5B). Apart from the two vaginal microbiota data sets, the best fit values of σ ranged from 0.1 to 0.3, indicating that changes in resource levels smaller than this magnitude will generate abundance changes that look like typical fluctuations. The best fit values of k varied between 0.5 and 1 across data sets, suggesting that the dynamics of microbial communities occur faster than or comparable to the typical sampling frequency of longitudinal studies. While it is unclear whether the internal time scales are faster than the sampling frequency for all of these communities, simulation results were robust to the dilution factor and threshold change defining ecological steady state (Figure 1—figure supplement 1), two main factors that affect the relationship between the internal and sampling time scales.

Inferences about intrinsic parameters of resource competition and external parameters of environmental fluctuations were also consistent with expectations for in vitro passaging of complex communities derived from humanized mice (Aranda-Díaz et al., 2022). The resulting time series statistics were best fit by the smallest value of σ among the data sets studied, indicating that the in vitro environment has relatively low noise across sampling times (as expected); the nonzero σ presumably arises from technical variations that result in effective noise in resource levels. The best fit value of M was larger than the reservoir size N, suggesting that there are many distinct resources in the complex medium used for passaging and consistent with the ability of more diverse inocula to support more diverse in vitro communities (Aranda-Díaz et al., 2022). The consistency of these results further supports the utility of our model.

Taken together, our model infers ensemble-level parameters of resource competition and external parameters of environmental fluctuations for several widely studied microbial communities that can inform future mechanistic studies.

Under a serial dilution scheme, an ecological steady state is reached when the dynamics in subsequent passages are identical, which is the case when all consumers are either extinct or have a growth ratio (the ratio of a consumer’s final and initial abundances within one passage) equal to the dilution factor D. Due to the slow path to extinction of some consumers, reaching an exact ecological steady state can require hundreds of passages, presumably more than realistically occurs between sampling times in the data sets examined here. Thus, we assumed instead that between sampling times the system only approximately reaches an ecological steady state, defined as the growth ratios of all species changing by less than a threshold between subsequent passages that was defined as a fraction of D. Throughout this study, D was set to 200 and the steady state threshold was 5%, under which a steady state was approximately reached in about 5 dilutions (Figure 1B). In this manner, our model assigns a well-defined state of consumer abundances to each resource environment while ensuring that only a reasonable amount of change occurs between sampling times. Note that in human gut microbiotas, abundances can change by more than 1000-fold between daily samplings (Figure 2B), indicating that at least 10 generations can occur between sampling times. The precise value of D did not affect time series statistics, and steady-state thresholds between 1% and 10% generated similar time series statistics (Figure 1—figure supplement 1). We therefore expect our results to be robust to the values of these two parameters. Simulations were carried out in Matlab, and all code is freely available online in Matlab and Python at https://bitbucket.org/kchuanglab/consumer-resource-model-for-microbiota-fluctuations/.

Saturation kinetics were implemented into the CR dynamics of Equation 1 asdXidt=Xi(∑j=1MRijYjYs+Yj),dYjdt=−YjYs+Yj(∑i=1NRijXi),

where Ys denotes the saturation constant. For simplicity, Ys was assumed to be equal for all resources, and set to an intermediate value of Ys=Yj,0/3 such that both saturated and linear kinetics could affect community dynamics. Other model details are the same as the original CR model.

The gLV model in Equation 3 was parametrized in two ways. The first parametrization, which we refer to as CR-converted gLV models, was motivated by the successful recapitulation of experimental time series statistics with our CR model. The CR model can be rewritten as a gLV model when resource consumption rates are similar to the mean value (Good et al., 2018). Under this assumption, the mapping is ri=2∑jRijYj,0 and Aij=1Rmax∑kRikRjkYk,0 . The converted interaction coefficients are negative and symmetric, and their magnitudes depend on the niche overlap between the interacting taxa. Since the resource levels Yj,0 are involved in this parametrization, fluctuations in Yj,0 across sampling times T translate into fluctuations in ri and Aij .

In the second parametrization, ri and Aij were randomly drawn from normal distributions with means and variances equal to those in the CR-converted gLV model. Aij were forced to be negative and symmetric.

The gLV models were initialized with equal relative abundances for all taxa, and simulated for a fixed amount of time such that a similar range of relative abundances was generated as in the CR model at approximate ecological steady state.

Raw 16S sequencing data from David et al., 2014; Song et al., 2020, were downloaded from the European Nucleotide Archive and the Sequence Read Archive, respectively, and ASVs were extracted using DADA2 (Callahan et al., 2016) with default parameters. OTUs or ASVs from other studies were downloaded and analyzed in their available form. All code for data processing is available in the repository listed above.