Metabolic complexity drives divergence in microbial communities

Abstract

Microbial communities are shaped by environmental metabolites, but the principles that govern whether different communities will converge or diverge in any given condition remain unknown, posing fundamental questions about the feasibility of microbiome engineering. To this end, we studied the longitudinal assembly dynamics of a set of natural microbial communities grown in laboratory conditions of increasing metabolic complexity. We found that different microbial communities tend to become similar to each other when grown in metabolically simple conditions, but diverge in composition as the metabolic complexity of the environment increases, a phenomenon we refer to as the divergence-complexity effect. A comparative analysis of these communities revealed that this divergence is driven by community diversity and by the assortment of specialist taxa capable of degrading complex metabolites. An ecological model of community dynamics indicates that the hierarchical structure of metabolism itself, where complex molecules are enzymatically degraded into progressively simpler ones which then participate in cross-feeding between community members, is necessary and sufficient to recapitulate our experimental observations. In addition to helping understand the role of the environment in community assembly, the divergence-complexity effect can provide insight into which environments support multiple community states, enabling the search for desired ecosystem functions towards microbiome engineering applications.

Introduction

Understanding how diverse microbial communities assemble is important for addressing open challenges in microbial ecology with applications that range from medicine^1,2 to climate change mitigation^3–5. Studies in natural^6,7 and laboratory^8–15 settings have investigated the reproducibility of assembly dynamics across a range of environmental conditions leading to seemingly contradictory results. Under certain conditions, microbial community assembly appears to be highly deterministic, as different communities are driven by strong environmental selection towards a specific steady state independent of their initial composition⁸. Under other conditions, however, environmental selection is weaker, resulting in highly variable assembly of communities with more dependence on their initial composition¹¹. Uncovering what properties govern this variability in microbial community assembly constitutes one of the fundamental questions of microbial ecology¹⁶ and is crucial for successful microbiome engineering, which aims to steer communities towards a desired structure in a given environment¹⁷.

Previous studies have explored how one community behaves across a range of conditions^9,10,18 or how communities behave under a single condition^11–15. However, systematically comparing the fates of multiple communities grown in multiple conditions remains an underexplored avenue for understanding microbial ecosystem dynamics. Culturing multiple communities in a given condition is necessary for understanding how closely those communities converge or diverge taxonomically from each other over time. Conversely, exploring these dynamics across many conditions is necessary for understanding how divergence depends on properties of the environment.

Here, we combined experimental measurements and computational modeling to investigate the interplay of initial composition and environmental selection in determining the extent to which the same set of distinct initial communities tend to resemble each other within environments that span a broad range of metabolic complexity. We followed the dynamic assembly of diverse microbial communities inoculated from different soil samples grown on carbon sources of increasing metabolic complexity. We then measured the divergence (beta-diversity) between communities within each condition. Since metabolic complexity could be measured by (i) the potential for a single substrate to give rise to new byproducts and (ii) the number of different substrates present in the environment¹⁸, we investigated single- and mixed-metabolite conditions, separately. While other studies have explored the relationship between metabolic complexity and (alpha) diversity within a community^9,18, we now investigate how metabolic complexity relates to the divergence between the same set of communities (beta diversity).

By tracking how closely the taxonomic structure of these communities resembled each other over time, we found that the effect of environmental selection on communities depended on the metabolic complexity of the environment itself. Specifically, different microbial communities diverged in their taxonomic composition across a gradient of increasingly complex metabolic conditions, suggesting that the forces dominating microbial community assembly shift from strong to weak environmental selection in increasingly complex conditions. We tested whether a commonly used ecological model for microbial ecosystem dynamics^8,19–21 could reproduce this effect and found that this occurs only upon incorporating a representation of the hierarchical structure of metabolite transformations (e.g. polysaccharides to oligosaccharides to monosaccharides)^14,22 and does not depend on how metabolite preferences are distributed across taxa. Our results point to an ecosystem organization principle that can help reconcile seemingly incompatible observations of divergence in different conditions and provide guidelines for which environments may be more susceptible to microbiome engineering projects.

Results

The divergence-complexity effect hypothesis

To assess the strength of environmental selection on community assembly, one would ideally compare how the trajectories of multiple distinct microbial communities diverge in taxonomic composition across a set of conditions. A key question we ask is whether distinct communities assembled in the same condition tend to become taxonomically similar and how the degree of similarity depends on the metabolic complexity of the environment. For example, we can imagine how different microbial communities that initially vary in taxonomic composition (Fig. 1a) may converge in composition over time when grown in one environment (strong environmental selection, Fig. 1b) while those same communities may diverge in another environment, arriving at alternative stable states (weak environmental selection, Fig. 1c). To quantify the degree to which different communities differ taxonomically from each other when grown in a given condition, we calculate the divergence (beta-diversity within each condition; see Methods and “Measuring divergence with the Aitchison distance metric” in Supplementary Information) in their compositions as they develop towards a steady state (Fig. 1d).

Fig. 1 |. Microbial communities may diverge in environments with increasing metabolic complexity.

a-d, Hypothesis of microbial community divergence in theoretical simple (b) and complex (c) metabolic conditions. Microbial communities A, B, and C are initially composed of different compositions of the same three microbial species (a; blue, red, and yellow). Over time, communities grown on a simple substrate (b) converge, while these same communities grown on a complex substrate (c) diverge. The lines in b and c show the trajectory of each community from the initial composition (circles) to final compositions (squares). d, Quantification of divergence (distance between communities in the same condition) at the final time point (trajectories arriving at squares shown in dashed circles above each bar for each condition) for hypothetical scenarios in a-c. e-g, Divergence observed in two independent experimental studies where microbial communities were sourced from soils or leaves and grown on glucose (e; a relatively simple metabolic environment from Goldford et al., N=11 communities) and communities were sourced from pitcher plants and grown on acidified cricket media (f; a more complex metabolic environment from Bittleston et al., N=10 communities). Each colored line in e and f represents the trajectory of a community’s composition over time in separately computed multidimensional scaling (MDS) projections. Circles indicate the initial community composition and squares indicate the final community composition. g, The divergence for each metabolic environment, calculated as the pairwise distances between all communities within a given condition at each time point. Each point is the mean pairwise distance within condition at each time point and shading represents the 95% confidence interval over all pairwise distances within each environment at each timepoint.

We initially identified existing data that could indicate whether and how community divergence would depend on environmental conditions. Specifically, we re-analyzed two independent studies that explored how a collection of diverse microbial communities assembled over time, but did so under very different conditions. When one study, Goldford et al.⁸, cultured communities in (simple) glucose media, communities converged (Fig. 1e). By contrast, when Bittleston et al.¹¹ cultured communities in (complex) acidified cricket media, they diverged (Fig. 1f). In both cases, the initial communities differed substantially from each other and then immediately became more similar; however, communities enriched on glucose ultimately converged significantly more, despite starting with greater variation in initial community composition (Fig. 1g). Based on the striking discrepancy in the degree of divergence across these two studies we formulated the hypothesis that divergence increases with the metabolic complexity of the provided resources (Fig. 1d), a relationship that we will refer to as the divergence-complexity effect.

Community divergence increases with metabolic complexity

To directly test the divergence-complexity effect, we designed an experiment to quantify the divergence of microbial communities grown in conditions of increasing metabolic complexity (Methods; Fig. 2a). To assess divergence, we sourced six microbial communities from forest soils, which are generally diverse and distinct from each other²³, even over small (centimeter) spatial scales⁷. Each microbial community was grown in nine different minimal media, each supplemented with equimolar carbon concentrations of at least one carbon source commonly found in soils²⁴: (1) citrate, (2) glucose, (3) cellobiose, (4) cellulose, (5) lignin, (6) citrate + glucose, (7) citrate + glucose + cellobiose, (8) citrate + glucose + cellobiose + cellulose, or (9) citrate + glucose + cellobiose + cellulose + lignin (Methods).

Fig. 2 |. Divergence of microbial communities increases in environments of increasing metabolic complexity.

a, Study design: microbial communities were extracted from six forest soils and were then grown in nine conditions (citrate, glucose, cellobiose, cellulose, lignin, citrate + glucose, citrate + glucose + cellobiose, citrate + glucose + cellobiose + cellulose, and citrate + glucose + cellobiose + cellulose + lignin). Communities were passaged ten times once every three days and sequenced on days 0, 3, 6, 9, 12, and 33 (N varies for each community, see Table 1). b-c, MDS projections of community trajectories over time in each single-metabolite condition (b) and mixed-metabolite condition (c). MDS was calculated on all samples together for ease of visually comparing trajectories between conditions. Circles indicate the initial community composition and squares indicate the final community composition. d, Divergence of communities within each condition over time from day 3 onwards. Initial communities are a distance of 58.1+/−3.5 (not shown for clarity). Single metabolite conditions are in blue, mixed conditions are in orange, and colors darken with complexity. Points on each line represent the mean divergence and the shaded region represents the 95% confidence interval for pairwise distances between all six communities within each condition. e, Distribution of divergence between all six communities at the final time point where divergence increases with metabolic complexity for single and mixed-metabolite conditions (same colors as d). Each point is a pairwise distance between two communities (N=15 per condition). Boxes are bound by the interquartile range, divided by the median, and whiskers extend to a maximum of 1.5 times the interquartile range. f, Divergence-complexity effect by condition type (slopes shown in e) for all time points (one-sided Wald Test on effect (slope) > 0, p-values shown in figure for each day and condition type). All divergence results for communities at day 33 are the mean distances between up to three replicates for each community (Extended Data Table 1).

In testing the divergence-complexity effect, we consider metabolic complexity to increase from citrate to lignin (in line with the number of metabolic byproducts expected from each metabolite¹⁸) enabling us to define a clear gradient of complexity separately for single metabolites, from condition (1) to (5), and mixed metabolites from condition (6) to (9). Lignin can be depolymerized extracellularly by a variety of enzymes into a diversity of polyphenolic compounds, which can be broken down into aromatic compounds, and eventually cleaved into respirable or fermentable byproducts^25–29. Cellulose, a polysaccharide composed of repeating glucose subunits, can be degraded by cellulases³⁰ into cellobiose (a disaccharide of a pair of glucose molecules) and glucose itself³¹. Citrate³², the simplest carbon source in our experiment, is an organic acid related to each of the metabolites in our study as a byproduct of respiration. Given that our culture conditions were likely not fully aerated, the above metabolites are expected to ultimately be used by the communities through a mixture of aerobic and anaerobic metabolism, not unlike what is experienced by soil microbial communties^33–36.

Each microcosm, containing one source community growing in one condition, was serially passaged ten times, in intervals of three days. 16S rRNA sequencing was performed and amplicon sequence variant (ASV) counts were generated for the initial soil inocula and for microcosm communities at days 3, 6, 9, 12, and 33 (excluding 48/384 samples, see Extended Data Table 1 and Methods). The divergence for each condition was then calculated by computing the distance between all pairs of communities within each condition for each day (Methods). For increased accuracy at our final time point, the divergence for day 33 communities was computed by averaging the distances between three replicates for each community (excluding one community, where only two replicates were available, see Extended Data Table 1).

Supporting our hypothesis of the divergence-complexity effect, we observed that divergence increased for both of our axes of metabolic complexity (single- and mixed-conditions; Fig. 2b–f). In accordance with previous studies (Fig. 1e–f), our source communities started differently from each other and initially converged (Fig. 2b–c) and stabilized once introduced to laboratory conditions (Extended Data Fig. 1, Fig. 2d). Within single-metabolite conditions, the communities converged strongly on simple metabolites, while they diverged to increasingly distinct states on the more complex metabolites (Fig. 2b, 2d–f). At the final time point, divergence occurred in a discrete fashion with similarly low levels between citrate and glucose, similarly high levels between cellulose and lignin, and an intermediate level of divergence on cellobiose (Fig. 2e) which may be related to where each metabolite is enzymatically processed (citrate and glucose: intracellularly, cellulose and lignin: extracellularly, and cellobiose: both ways³⁷). Similar to single-metabolite conditions, community divergence increased from the least (citrate + glucose) to the most diverse (all metabolites) mixed-metabolite conditions (Fig. 2c–f). For single- and mixed-metabolite conditions, the divergence-complexity effect was detectable at each sampled time point (Fig. 2f; Extended Data Fig. 2), suggesting that the effect occurred rapidly (as early as day 3) and was sustained irrespective of possible evolutionary changes. While evolutionary adaptation to utilizing new nutrients tends to occur at much longer timescales^38–40, we cannot rule out the role of strain evolution in our longitudinal analysis⁴¹.

Interestingly, the effect was stronger in single-metabolite conditions than in mixed-metabolite conditions (Fig. 2e–f). One possible explanation for this trend includes priority effects⁴², where assembly dynamics are dependent on the order in which different metabolites become available through trophic interactions¹⁰. Another possibility is that divergence was sensitive to the concentration of complex metabolites⁴³. Since the total carbon concentration was kept constant across all conditions, the concentration of complex metabolites was highest in single-metabolite conditions and decreased incrementally for each subsequently complex mixed-metabolite condition (Methods).

Since it is common for many taxa to engage in functionally redundant actitivies^44,45 and since our data is generated from amplicon sequencing, we cannot conclude whether these taxonomically divergent communities are utilizing similarly divergent metabolic pathways⁴⁶. However, when we re-computed divergence at the Family taxonomic level (Extended Data Fig. 3–4) where bacteria often differ in metabolic functions¹⁵, we still observed the divergence-complexity effect, which suggests that our communities, that assembled to distinct taxonomic compositions, may have also engaged in distinct metabolic activities.

Divergence, which was high across different source communities in complex conditions, was significantly lower across replicates from the same source community. Replicate microcosms assembled from the same source converge to similar final states after 33 days (Extended Data Fig. 5). Conversely, communities sourced from different locations, even if they displayed similar initial composition, did not necessarily converge (Fig. 2b–c). This suggests that while assembly dynamics are indeed complex (initially similar communities can diverge), they are reproducible (replicate microcosms converge). While previous studies had reported that stochasticity can be responsible for multistability, this similarity across replicate experiments suggests that stochasticity is not the main driver of divergence in our communities¹⁵. Our observations are consistent with simulation work that suggests that multistability can take place without stochastic fluctuations in environments containing multiple essential nutrients⁴⁷. Ultimately, the community assembly trajectories we observed in our experiment show that complex environments can support a greater number of discrete states than simple environments, but that these discrete environments may be achievable only from distinct initial conditions.

Diversity correlates with divergence implicating specialists

In order to gain a deeper understanding of the divergence-complexity effect, we investigated how alpha diversity within each individual community correlates with divergence across communities. In particular, two separate principles could jointly give rise to the divergence-complexity effect. The first principle, “metabolic complexity begets diversity”, where community diversity increases with increasing metabolic complexity, has been experimentally documented in both natural and synthetic communities^9,18. A yet unexplored second principle, that we investigate in our study, “diversity begets divergence”, could result from the expectation that more diverse communities generally diverge more from one another (see “Divergence is sensitive to richness and evenness” in Supplementary Information; Fig. 4d). In general, the amount of divergence we observe is greater than we expect from richness alone, suggesting that differing distribution of taxa (skewness) between communities substantially contributes to the divergence-complexity effect (Supplementary Fig. S1). If metabolic complexity were to yield diversity and diversity were to yield divergence, we would expect higher divergence in increasingly complex conditions, contributing to the divergence-complexity effect.

Fig. 4 |. Endemic taxa are enriched and unevenly distributed in complex conditions.

a, The distribution of taxon condition-specificity per condition for day 33. The number of taxa (N) for each condition is indicated in the figure. Condition-specificity is calculated as the fraction of occurrences of a taxon that is attributed to a particular condition, such that a specificity of 1 means that taxon occurs in only one condition (a specialist). Boxes are bound by the interquartile range, divided by the median, and whiskers extend to a maximum of 1.5 times the interquartile range. Diamond symbols indicate outliers passed 1.5x beyond the interquartile range. b, The number of specialists per condition (taxa with condition-specificity=1). c, Taxon occurrence by number of conditions and number of source communities. ASVs found in fewer conditions are less evenly distributed across source communities (found in fewer source communities) and taxa found in more conditions are more evenly distributed across source communities. d, Two hypotheses for single metabolite conditions following from a and b, where H2 is supported and H1 is not. H1: more complex conditions are enriched for specialists and when those taxa are evenly distributed across source communities, it results in relatively similar divergence for complex and simple metabolic conditions. H2: when more complex conditions are enriched for specialists and these taxa are less evenly distributed across communities, more complex conditions result in greater divergence.

Consistent with these expectations, we observed a strong linear relationship between diversity and divergence, which strengthened over time, indicating that specific changes in community assembly drive the rise of divergence. The slope of the diversity-divergence relationship increased over time (black line Fig. 3a; Fig. 3b; Extended Data Fig. 6), despite the fact that diversity itself, on average, decreased (red line Fig. 3a). In other words, over time, the same degree of divergence is maintained by communities with reduced diversity. For divergence to remain relatively stable (Fig. 2d) while diversity decreases, taxa endemic (i.e. specific) to each community must persist while a set of species shared across communities universally go extinct within each condition. One possible explanation for this trend is that these persistent taxa are metabolic specialists, which produce enzymes that target specific biochemical bonds⁴⁸. In conjunction with our previous observations that communities diverge more in complex conditions, we hypothesized that specialists which target complex metabolites were less evenly distributed across communities than taxa that specialize on simpler metabolites, driving the divergence-complexity effect.

Fig. 3 |. Divergence dynamically correlates with diversity.

a, The slope of the relationship between community alpha diversity and divergence (black) and the mean community alpha diversity (red) over time. Shaded areas around each regression line represents the 95% confidence interval. b, The data underlying the relationship in a over time. Each point is the diversity of a community in a condition (x-axis; same colors per condition as Figure 2) and the mean divergence of that community from all others within a condition (y-axis). All divergence results for communities at day 33 are the mean distances between up to three replicates for each community (Extended Data Table 1). The coefficient of determination (R²) for each day is reported from linear regression of mean divergence on diversity. While diversity is expected to increase with metabolic complexity, it is not clear if different communities will increase in diversity in the same ways (with coordinated changes to the same taxa). See Supplementary section “Divergence is sensitive to richness and evenness” for more details about the relationship between divergence and diversity.

Specialists are more endemic in complex conditions

To investigate the role of specialists in the divergence-complexity effect, we explored the distribution of taxa across experimental conditions and source communities. If specialists drive the diversity-complexity effect, we would expect to see that specialists are increasingly endemic, or unevenly distributed across source communities, in more complex conditions.

To quantify the degree of specialization, we computed a condition-specificity metric for each taxon (ASV) in each condition, and then assessed whether specialization and endemism depended on metabolic complexity. We defined condition-specificity for each taxon and condition as the fraction of source communities in which that taxon was found at the final sampling time point for that given condition. In particular, if a taxon only occurs in one condition, its condition-specificity is 1 and will be referred to as a “specialist”. In accordance with our expectations, we observed that more complex conditions (particularly single-metabolite ones) had greater condition-specificity (Fig. 4a) and more specialists (Fig. 4b). Given the broad phylogenetic distribution of traits associated with degrading and utilizing the byproducts of cellulose⁴⁹ and lignin^25–27, these condition-specific taxa may be performing redundant metabolic functions in their respective communities.

These results alone are encouraging, but are not sufficient for linking specialists to the divergence-complexity effect, which would additionally require specialists to differ between communities in the same condition. While we observed an enrichment of condition-specific taxa in complex conditions, it hypothetically could be the case that these same taxa were found across all source communities, in which case communities in complex conditions would not necessarily diverge substantially more than those in simple conditions (H1 Fig. 4d; Supplementary Fig. S1). However, when we count the occurrence of each taxon in each condition and source community, we find that condition-specific taxa are also source community-specific (endemic; Fig. 4c). As a result, taxa that specialize on complex metabolites are less evenly distributed across communities than taxa that specialize on simpler metabolites and are therefore heavily implicated in mediating the divergence-complexity effect (H2 Fig. 4d; Supplementary Fig. S1).

Trophic resource transformations reproduce divergence

To better understand what properties of microorganisms and their environment are necessary for the diversity-complexity effect, we performed a series of simulations with microbial consumer-resource models (CRMs; Methods)¹⁹. In particular, we wanted to corroborate our hypothesized mechanism of the diversity-complexity effect, that divergence correlates with metabolic complexity and emerges from endemism of specialist taxa.

CRMs are dynamical ecological models where consumers are defined by the set of resources they prefer (consumer preferences) and resources are transformed by consumers into other resources following consumption (resource transformations). Overlapping consumer preferences give rise to competition, while the exchange of transformed products can generate cross-feeding interactions¹⁹. Cross-feeding can take many different shapes in natural communities, including adaptive cooperation among auxotrophs⁵⁰, secretion of central carbon metabolism intermediates^51–56 or fermentation products⁵⁷, extracellular degradation of complex macromolecules whose byproducts are shared with community members⁵⁸, or stress-induced cross-feeding⁵⁹. While cross-feeding in CRMs is often assumed to reflect exudation of intracellular intermediates¹⁹, CRMs can still be interpreted as approximating a broader set of mechanisms, including those mentioned above (see “Generality of cross-feeding mechanism in our consumer-resource model” in Supplementary Information).

In implementing a CRM, we sought to investigate how the divergence-complexity effect depends on consumer preferences and resource transformations. These consumers and resources are typically arranged in a trophic structure, where taxa specialize in the hierarchical consumption of environmentally available metabolites and cross-feed the resulting (simpler) byproducts to taxa at subsequently lower trophic levels^10,22. Trophic interactions are implicated in explaining the immense diversity of microbial communities since only certain organisms can break down the most complex resources in a given environment²² and, on average, half of a microbial community’s members are auxotrophic and require a cooperative interaction with another organism for essential nutrients⁵⁰.

Taxonomic structure in CRMs can be represented by specifying what resources each consumer prefers (Extended Data Fig. 7c–d). In our models, we defined trophic consumer preferences with “families” of specialists, or modules of consumers that have similar resource preferences (specialization)¹⁰, and generalists, which have a broad distribution of preferences (Extended Data Fig. 7c). In line with our experimental observations (Fig. 4), we set the number of different specialists to be proportional to the complexity of the resource type they prefer. We also generated random controls where consumers preferences were assigned randomly (Extended Data Fig. 7d). The use of such random structure has been used successfully to reproduce certain ecological patterns, suggesting that community assembly can be well approximated by random, dense, complex interactions between microbes and their environment^20,60.

Metabolic structure can be represented by assuming that upon metabolization, resources of a given type transform into resources of another specific type (Extended Data Fig. 7a–b. In our models, we defined trophic resource transformations such that complex resources transformed into simpler ones in a hierarchical fashion²⁰ (Extended Data Fig. 7a). Similar to random consumer preferences, we defined random resource transformations where resources of one type could transform into each other randomly (Extended Data Fig. 7b).

To investigate the role of taxonomic and metabolic structure in community divergence, we simulated communities growing in increasingly complex conditions. We performed these simulations (Fig. 5) using four different CRM configurations (Extended Data Fig. 7): trophic consumer preferences and trophic resource transformations (fully structured; Fig. 5a), random preferences and trophic transformations (resource structured; Fig. 5b), trophic preferences and random transformations (consumer structured; Fig. 5c), and random preferences and transformations (fully random; Fig. 5d). Similar to our experimental design, all four model configurations were initialized with six source communities and seven conditions (four single and three mixed resource conditions; Extended Data Fig. 7e–f) and growth dynamics were simulated until reaching a steady state. To understand whether the divergence-complexity effect could emerge solely from the ecological forces encoded in our model (consumer preferences and resource transformations), we assumed physiological parameters, such as rates of consumer growth, consumer maintenance, resource utilization, resource energy density, and leakage, to be uniform across all consumers and resources^19,20.

Fig. 5 |. Trophic resource transformations reproduce divergence-complexity effect with consumer-resource model simulations.

The distribution of divergence for six communities simulated with consumer-resource models with and without trophic structure in resource transformations and consumer preferences. Mimicking our experiment, community growth was simulated in single metabolite (blue) and mixed metabolite (orange) conditions of increasing complexity (from lighter to darker). a-d, Divergence for communities simulated with trophic resource transformations and consumer preferences (fully structured; a), trophic resource transformations and random consumer preferences (resource structured; b), random resource transformations and trophic consumer preferences (consumer structured; c), and random resource transformations and consumer preferences (fully random; d). Each box describes the distribution of pairwise distances between all six simulated communities (N=15 per box). Boxes are bound by the interquartile range, divided by the median, and whiskers extend to a maximum of 1.5 times the interquartile range. e-f, The effect of metabolic complexity on divergence for single and mixed metabolite conditions with trophic resource transformations (e) and random transformations (one-sided Wald Test f, p-values shown in figure for each parameterization and condition type). g-h, Using the fully structured configuration, the relationship between diversity and mean divergence where the shaded region represents the 95% confidence interval (g) and the relationship between occupancy in conditions and number of source communities (h).

Surprisingly, our simulations showed that trophic structure of the resource transformations alone was necessary and sufficient to reproduce the divergence-complexity effect in both single- and mixed-resource conditions (Fig. 5a–f). Even when consumer preferences were random, the divergence-complexity effect was still observed as long as resource transformations were structured (Fig. 5a,b,e). However whenever resource transformations were random, all communities diverged equally, irrespective of metabolic complexity, and thus there was no divergence-complexity effect (Fig. 5c,d,f).

In addition to reproducing the divergence-complexity effect, our model recovered, as emergent properties, further non-trivial trends detected in our experiment. For example, in model configurations with trophic resource transformations, the divergence-complexity effect is greater for single resource conditions than mixed ones (Fig. 5e), as observed experimentally (Fig. 2f). Additionally, the maximum divergence in single resource conditions exceeds that of mixed conditions (Fig. 5a–b; Fig. 2e). These model configurations also reproduced our downstream analyses, such as the correlation between divergence and diversity (Fig. 5g; Fig. 3b) and the tendency for specialists to be endemic (Fig. 5h; Fig. 4d). The reproduction of all of these patterns with physiologically neutral consumers (uniform physiological parameters) and resources implicates the trophic metabolic structure in resource transformations as the driving mechanism of the divergence-complexity effect.

Discussion

Compelled by recent experiments which found that microbial community diversity increases with metabolic complexity^9,18, we sought to reconcile contradictory interpretations of whether microbial communities tend to converge⁸ or diverge¹¹ in the same conditions. By jointly revisiting these two propositions, we hypothesized and provided support for the divergence-complexity effect as a quantitative ecological principle which resolves this contradiction. While previous work explored community assembly by modulating the complexity of metabolic conditions^9,10,18 or the variability of source communities^8,11,12, the divergence-complexity effect could be observed only by systematically varying both, i.e. analyzing multiple source communities under increasingly complex conditions. We found that divergence correlates strongly with diversity, which is driven by an enrichment of specialists in complex conditions. We concluded our analysis by reproducing these results using consumer-resource model simulations, which provide insights into the potential ecological mechanisms of the divergence-complexity effect.

While our experimental results are robust and reproducible, they necessarily rely on specific design constraints and assumptions. Experimental choices that could be revisited in future studies include the passaging time, chosen here to be three days, as used in other microbial community assembly studies with complex metabolites¹¹; the selection of metabolites, which constitute a representative, but oversimplified version of the metabolic complexity of soil environments; our plate shaking rate (Methods), which may have affected oxygen availability⁶¹ and the balance of aerobic and anaerobic metabolism; and the focus on taxonomic divergence (through 16S amplicon sequencing) rather than functional divergence, which would require a comprehensive profiling of microbial functions with metagenomics, metatranscriptomics, or metabolomics. Furthermore, future research could explore which environmental parameters, in addition to metabolic complexity, may drive divergence^62,63.

The most surprising result from our simulations was how structured resource transformations (where complex metabolites are progressively degraded into simpler ones), but not consumer preferences, were required for reproducing the divergence-complexity effect (Fig. 5). A possible interpretation of this result is that microbial community assembly and dynamics are strongly dependent on the actual structured architecture of metabolism, which differs substantially from a network of random transformations^20,60. To test the generality of the divergence-complexity, we tried to minimize the use of data-derived parameters in the model. One exception is the parameterization of structure in consumer preferences based on our observed distribution of taxa, which is in any case ultimately shown to be unnecessary to yield the divergence-complexity effect. Conversely, we cannot rule out that additional or alternative data inputs into the model could affect its performance. For example, our consumer-resource models lacked the encoding of certain processes that are known to affect microbial community assembly such as nuanced cross-feeding mechanisms (Supplementary Information), diffusion⁶⁴, transcriptional regulation⁶⁵, tradeoffs between growth rate and enzyme production^66,67, evolution⁴¹, and antibiotics⁶⁸. Explicit measurement of these processes in future studies or inclusion with other ecological models⁶⁹ could help to reveal further mechanistic insights into the diversity-complexity effect. However, the fact that our current model captures so many of our experimental observations lends confidence to the dominant role that the architecture of metabolism plays in community structure, corroborating previous reports¹⁸.

Importantly, the divergence-complexity effect has direct implications for the engineering of microbial communities towards any target, suggesting that metabolically complex environments may be more susceptible to microbiome engineering than simple ones. Potential targets for microbiome engineering include correcting the dysbiosis in the human gut² and increasing the carbon stabilization capacity of soils⁵, among many other microbially-regulated traits. The consequences of the divergence-complexity effect are encouraging for efforts along these lines, since complex environments may be more likely to support an alternative community that is equally stable as the original one, but with potentially increased expression of a trait of interest. Approaches such as directed evolution, where a set of microbial communities undergoes iterative rounds of perturbation and artificial selection in order to assemble high-performing communities⁷⁰, offer an ideal strategy for exploring the different alternative states that a complex environment can support. Future research is required to understand how, in light of functional redundancy, the divergence in taxonomic composition that we observe relates to divergence in functional composition, since modifying functional activity is commonly the goal of microbiome engineering efforts. Ultimately, we envisage that the awareness of the divergence-complexity effect may help microbial ecologists reframe the role of environmental selection in microbial community assembly and enable further research into the engineering of complex microbially-regulated environments.

Methods

Media preparation

Eleven different media were generated at equimolar (50mM) concentrations of carbon (C) in increasing levels of complexity. Stocks of citrate, glucose, cellobiose, cellulose, and lignin were generated at 1 mol C/L (1M C) and then sterilized. Citrate, glucose and cellobiose stocks were sterilized through 0.2um RapidFlow filters while cellulose and lignin stocks, whose particle sizes were too large for filters, were autoclaved. C source stocks were then mixed with M9 minimal media (5x M9 salts (BD Difco), 1M MgSO₄, 1M CaCl₂, 1000x trace minerals)⁷¹ to form the following nine conditions, each made at a final concentration of 50mM C and with equal ratios of each C source (Sigma-Aldrich) for each condition: citrate, glucose, cellobiose, cellulose, lignin, citrate + glucose, citrate + glucose + cellobiose, citrate + glucose + cellobiose + cellulose, and citrate + glucose + cellobiose + cellulose + lignin. All media were stored in glass bottles, wrapped in foil, and stored at 4°C.

Sample collection and microbial community extraction

On October 27, 2022 about half a pound of organic horizon soil (5–10cm deep) was collected from six sites at Harvard Forest in Petersham, MA, USA following sampling approval from Harvard Forest staff. Two were pine dominated, two were hardwood dominated, and two were mixed. Samples were collected 15m from the forest edge and kept on ice until transported back to the laboratory the same day. Fresh soils were sieved through a 2mm mesh and then stored at 4°C. On November 21, 2022, 20g of each sieved soil was individually combined with 100mL of sodium pyrophosphate to separate cells from soils⁷² and was blended for three cycles of 10 seconds at ~22,000 RPM and then off for 10 seconds, and then 25mL of the resulting slurry was transferred to a centrifuge tube. The blender was washed between each sample by blending in 500mL of diluted bleach. Following the blending of all soils, each slurry was centrifuged for 10 minutes at 20,000xg, resuspended in 30mL of PBS, and rocked on an orbital shaker for 1 hour¹⁸ at 4°C. After rocking, samples were allowed to settle for 5 minutes and then passed through a 100um cell straining filter. Optical density (OD) measurements were performed at 600nm at a 1:20 dilution, 500uL of each sample was stored at −80°C in 20% glycerol, and the remaining volume from each sample used for inoculating experimental plates.

Experimental culturing

Community extracts were added to 96-deep well plates in triplicate with all media combinations (3 replicates of each source community in each condition), generating a total of 162 microcosms (9 media combinations × 6 source communities × 3 replicates). Cycloheximide, an antifungal agent, was added to each well at 200ug/mL⁷³ to reach a final OD of 0.1 and volume of 400uL per well. Plates were then stored in an incubator at 25°C under constant shaking at 200RPM. Communities were passaged every 72 hours into fresh media at a 1:20 dilution (without cycloheximide) for a final volume of 400uL, OD₆₀₀ was measured, and the remaining volume was stored at −80°C for DNA extraction.

DNA extraction and sequencing

DNA extraction for the six initial forest soil communities was performed using the PowerSoil DNA extraction (QIAGEN). After adding lysis buffer, samples underwent three cycles of freezing in liquid nitrogen, warming at 55°F in a water bath, and bead-beating for 1 minute (PowerLyzer, MoBio), then following the provided protocol for the remainder of the extraction. DNA was extracted from an additional 330 lab cultured samples on different days for each forest site and condition on days 3, 6, 9, 12, and 33, where available (Extended Data Table 1). DNA extraction from lab cultured samples was performed using the PureLink Pro 96 Genomic DNA Kit (ThermoFisher) following the provided protocol except for extending all lysis incubation periods to two hours. DNA extracts were sent to Quintara Biosciences for library preparation and 16S amplicon sequencing using V4 primers 515F (GTGYCAGCMGCCGCGGTAA) and 806R (GGACTACNVGGGTWTCTAAT) on a single Illumina MiSeq run.

Amplicon sequence processing

We received raw sequencing data for our study from Quintara Biosciences and downloaded raw sequencing data from Goldford et al.⁸ (SRP144982) and Bittleston et al.¹¹ (SRP218147) from NCBI. All raw 16S sequencing data for each study was separately processed using BU16S (https://github.com/Boston-University-Microbiome-Initiative/BU16s), a QIIME2⁷⁴ pipeline customized to run on Boston University’s Shared Computing Cluster. Briefly, BU16S first trims primers and filters out reads of less than 50 base pairs using cutadapt⁷⁵, then obtains ASVs using dada2⁷⁶, and finally classifies ASVs with 95% or greater sequence identity to the SILVA_132_99 database with VSEARCH⁷⁷.

Data analysis

All data analysis was performed in Python version 3.8.11. Pairwise distances between samples were computed with the Aitchison distance because it accounts for the compositional nature of sequencing data, unlike common distance metrics, such as Bray-Curtis, Jenson-Shannon Divergence, and Unifrac^78,79. The Aitchison distance, A, between two compositions x and y is:

$$\mathrm{A}(\mathbf{x},\mathbf{y})=\left|\right|\mathrm{c}\mathrm{l}\mathrm{r}\left(\mathbf{x}\right)-\mathrm{c}\mathrm{l}\mathrm{r}\left(\mathbf{y}\right)\left|\right|$$

$$\mathrm{c}\mathrm{l}\mathrm{r}\left(\mathbf{x}\right)=\mathrm{l}\mathrm{o}\mathrm{g}(\mathbf{x}/\mathrm{G}(\mathbf{x}\left)\right)$$

$$\mathrm{G}(\mathrm{x})=(\mathrm{x}1,\dots \mathrm{x}\mathrm{n}{)}^{1/\mathrm{n}}$$

Where clr is the center-log ratio transform and G is the geometric mean. Divergence was computed by calculating the Aitchison distance between all pairs of samples within a condition at each time point (see Supplementary section “Measuring divergence with the Aitchison distance metric” for more details). The divergence for samples at day 33, where we have up to three replicates for each community, was reported as the mean pairwise distance between all replicates in all analyses. The divergence-complexity effect was calculated for each condition type (single- and mixed-metabolite conditions) and time point as the slope of a linear least-square regression with metabolic complexity as the independent variable and divergence between all community pairs as the dependent variable. In these linear models, the metabolic complexity of each condition was treated ordinally, increasing from citrate to lignin for single-metabolite conditions and from citrate+glucose to all metabolites for the mixed-metabolite conditions. The significance of the divergence-complexity effect for condition and time point was computed with a one-sided Wald Test with a null hypothesis of a slope of zero. The mean divergence for each sample in a given condition was computed as the mean pairwise distance from each sample to all other samples in that condition.

Dimensionality reduction of pairwise distances was performed using multidimensional scaling (MDS) in scikit-learn⁸⁰. MDS was computed separately for samples from Goldford et al. and Bittleston et al., while for our data MDS was computed jointly on all samples to allow for ease of comparability when viewing community trajectories in separate conditions.

Alpha diversity was computed by first rarefying (subsampling) all samples to 5028 reads and dropping twelve samples below this sequencing depth from subsequent alpha diversity analyses. The Shannon Diversity Index was calculated as $-\mathrm{\Sigma }\mathbf{x}\mathrm{l}\mathrm{n}\mathbf{x}$ and the ecological richness was calculated as Σx>0 for each sample composition, x.

Condition specificity was calculated for each ASV by calculating the fraction of times each ASV was present in each condition. ASVs with a condition specificity of 1 were considered “specialists” since they were only found to occur in a single condition.

Consumer-resource models

We simulated the growth of 168 microcosms (6 communities × 7 conditions × 4 configurations) using microbial consumer-resource models (CRM). All model parameters are discussed below and summarized in Extended Data Table 2. With microbial CRMs, the dynamics of species and resources can be modeled with the following equations:

$$\frac{d{N}_i}{dt}=N_i\left({\sum }_{\alpha }\left(1-l\right)c_{i,\alpha }{R}_{\alpha }-m\right)$$

$$\frac{d{R}_{\alpha }}{dt}=\left(R_{\alpha }^0-R_{\alpha }\right)-{\sum }_j{N}_j{c}_{j,\alpha }{R}_{\alpha }+{\sum }_{j,\beta }{N}_j{c}_{j,\beta }{R}_{\beta }{D}_{\alpha ,\beta }l$$

Where $N_i$ is the abundance of species $i$, $R_{\alpha }$ is the concentration of resource $\alpha$, $R_{\alpha }^0$ is the resource supply concentration, $l$ is the leakage fraction i.e. how much each resource is “leaked” (how much of $\alpha$ is converted into $\beta$, where the rest is converted into biomass), $m$ is the consumer maintenance cost, $c_{i,\alpha }$ is the consumer preference matrix, and $D_{\alpha ,\beta }$ is the resource transformation matrix describing the rate that $\beta$ turns into $\alpha$ following consumption¹⁹. We fixed the leakage ($l$=0.8), maintenance ($m$=1), and uptake rates ($c_{i,\alpha }$={0, 1}) for all consumers resulting in an “ecologically neutral” model.

In order to study the impacts of trophic structure on divergence, we explored four different CRM configurations that varied in whether or not resource transformations or consumer preferences were trophically structured or random (Extended Data Fig. 6). Resources were defined by establishing a resource pool of four resource types, T0, T1, T2, and T3, where each type consisted of 80, 60, 40, and 20 resources, respectively. Trophic resource transformations were parameterized by defining resources of one type to subsequently transform into resources of another type in a unidirectional fashion, with some self-renewal (Extended Data Fig. 6a). For example, T0-type resources mostly transform into T1-type resources and some T0-type resources (Extended Data Fig. 6a). Random resource transformations were defined by allowing each resource transform into any other resource with uniform probability (Extended Data Fig. 6b). Transformation profiles for each resource in both configurations were sampled from Dirichlet distributions.

Consumers were defined by establishing a metacommunity of four “families”, F0, F1, F2, and G, where each type consisted of 500, 300, 100, and 100 consumers, respectively. Trophic consumer preferences were defined by allowing consumers of each family to utilize a total of 35 sampled resources from their associated type (i.e. F0 consumers could utilize T0 resources) and a common resource type. The skewed distribution of consumer family size was chosen to model our experimental results where the number of specialists correlated with metabolite complexity. Consumers belonging to the G (generalist) family could consume resources of any type (Extended Data Fig. 7c). Random consumer preferences were defined by allowing each consumer to utilize 35 random resources from any type (Extended Data Fig. 7d). Code for sampling trophic resource transformations and trophic consumer preferences can be found at: https://github.com/michaelsilverstein/ms_tools/blob/main/ms_tools/crm.py.

Initial conditions and source communities were defined to mimic our experimental design. Seven conditions were defined by sampling 20 resources each with an initial abundance of 50 from each resource type for single-metabolite conditions (T0, T1, T2, and T3) and from mixtures of resource types for mixed-metabolite conditions (T3+T2, T3+T2+T1, T3+T2+T1+T0; Extended Data Fig. 7e). Six source communities were defined by sampling 200 consumers from the metacommunity each with an initial abundance of 1 (Extended Data Fig. 7f).

The dynamics of each source community was then simulated in each condition using all four parameter configurations (trophic transformations and preferences, trophic transformations and random preferences, random transformations and trophic preferences, and random transformations and preferences) and the divergence was computed for each condition and configuration. Simulations of community assembly were performed by passing model parameters (D matrix, c matrix, and initial conditions) to the Community Simulator package²¹, which provides utility functions for constructing and solving the system of ordinary differential equations. In order to appropriately compare our simulation results, which simulates actual abundances of each consumer, to our experimental results, which reports relative abundance of each ASV, we rescaled the abundance of all communities to the same range to simulate the process of sequencing. Divergence was then calculated on the rescaled simulated community composition profiles in the same way as with our experimental data (using the Aitchison distance).

Extended Data

Extended Data Fig. 1 |. Microbial communities stabilize.

Each point is the mean distance to a community’s previous timepoint within each condition and the shaded region represents the 95% confidence interval. On day 3, communities are substantially different from their initial state and then continue to change in composition, but by approximately the same amount.

Extended Data Fig. 2 |. Divergence of a subset of communities which are available for all days.

A reproduction of Fig. 1d–f, but excluding all samples from communities HF1P and HF3H, which are missing data for days 3, 6, and 9 (Extended Data Table 1). a, Divergence of communities within each condition over time from day 3 onwards. Points on each line represent the mean divergence and the shaded region represents the 95% confidence interval for pairwise distances between all four included communities within each condition. b, Distribution of divergence for the final time point. Boxes are bound by the interquartile range, divided by the median, and whiskers extend to a maximum of 1.5 times the interquartile range. c, Divergence-complexity effect by condition type for all time points (one-sided Wald Test on effect (slope) > 0, p-values shown in figure for each day and condition type.

Extended Data Fig. 3 |. Divergence-complexity effect at Family level.

A reproduction of divergence-complexity effect calculation in Fig. 2f except computing divergence at the Family level instead of at the ASV level (one-sided Wald Test on effect (slope) > 0, p-values shown in figure for each day and condition type). effect (slope) > 0. Unlike our results at the ASV level, the divergence-complexity effect for our mixed-metabolite conditions at the Family level was only significant for day 33. The lack of significance at other points may be related to their reduced sample size (for days 3, 6, and 9; see Extended Data Table 1), but may in principle reflect additional effects emerging upon taxonomic coarse graining.

Extended Data Fig. 4 |. Family composition over time.

Each subplot shows the Family composition over time for a microcosm. Colors for each Family are defined by Phylum where blue represents Actinobacteria, yellow represents Bacteroidetes, green represents Firmicutes, and red represents Proteobacteria. Each column of subplots pertains to one source community and each row pertains to a condition (increasing in complexity from top to bottom, with single- and then mixed-metabolite conditions).

Extended Data Fig. 5 |. Microcosm replicates cluster.

a, The final state at day 33 of two replicates for community HF1P and three replicates for all other communities (see Extended Data Table 1) projected separately for each condition using MDS. b, The distribution of distances across communities (N=1,066 pairwise distances) and between replicates within the same community (N=142 pairwise distances). Each violin outlines the kernel density estimate and contains a box which is bound by the interquartile range with an open circle at the media and whiskers that extend up to 1.5 times the interquartile range. Communities are colored by their source and squares represent the replicate used in the main text. Independent of condition, communities from the same replicate are more similar to each other than communities from separate replicates.

Extended Data Fig. 6 |. Divergence and diversity on a subset of communities which are available for all days.

A reproduction of Fig. 3, but excluding all samples from communities HF1P and HF3H, which are missing data for days 3, 6, and 9 (Extended Data Table 1). a, The slope of the relationship between community alpha diversity and divergence (red) and the mean community alpha diversity (gray) over time. Shaded areas around each regression line represents the 95% confidence interval. b, The data underlying the relationship in a over time. Each point is the diversity of a community in a condition (x-axis; same colors per condition as Figure 2) and the mean divergence of that community from all others within a condition (y-axis).

Extended Data Fig. 7 |. Consumer resource parameterizations.

D Matrices for trophic (a) and random (b) resource transformations. Each matrix column defines which resources are generated from a given input resource. a, Trophic resource transformations were parameterized by defining resource types (T0, T1, T2, and T3; red lines) which mostly transform into resources of the subsequent type with some self-renewal. b, Random resource transformations were parameterized by defining any resource to transform into any other resource with uniform probability. The colorbar indicates log₁₀ resource transformation rate for a and b. C matrices for trophic (c) and random (d) consumer preferences. Each matrix row defines which resources can be utilized by each consumer. c, Trophic consumer preferences were parameterized by defining consumer families (F0, F1, F2, and G; blue lines) which consume a total of 35 resources of their associated type (for example, F0 consumers utilize T0 resources) and a common resource, T3. G consumers (generalists) consume resources of any type. D, Random consumer preferences were parameterized by defining consumers to utilize any resource with uniform probability. e, Seven initial environmental conditions were defined with 20 resources sampled from (1) T0, (2) T1, (3) T2, (4) T3, (5) T3+T2, (6) T3+T2+T1, and (7) T3+T2+T1+T0. f, Six source communities were defined by sampling 200 consumers.

Extended Data Table 1 |. Availability of each sample.

When available, one sample was sequenced from each microcosm for days 0, 3, 6, 9, and 12. Three samples were sequenced for day 33. Bolded text indicates unsequenced conditions.

Day	Condition	HF1P	HF2P	HF3H	HF3E	HF1M	HF2M	Total
0		1/1	1/1	1/1	1/1	1/1	1/1	6
3	Citrate	0/1	1/1	0/1	1/1	1/1	1/1	4
	Glucose	0/1	1/1	0/1	1/1	1/1	1/1	4
	Cellobiose	0/1	1/1	0/1	1/1	1/1	1/1	4
	Cellulose	0/1	1/1	0/1	1/1	1/1	1/1	4
	Lignin	0/1	1/1	0/1	1/1	1/1	1/1	4
	C+G	0/1	1/1	0/1	1/1	1/1	1/1	4
	C+G+C	0/1	1/1	0/1	1/1	1/1	1/1	4
	C+G+C+C	0/1	1/1	0/1	1/1	1/1	1/1	4
	C+G+C+C+L	0/1	1/1	0/1	1/1	1/1	1/1	4
6	Citrate	0/1	1/1	1/1	1/1	0/1	1/1	4
	Glucose	0/1	1/1	1/1	1/1	1/1	1/1	5
	Cellobiose	0/1	1/1	1/1	1/1	1/1	1/1	5
	Cellulose	0/1	1/1	1/1	1/1	0/1	1/1	4
	Lignin	0/1	1/1	1/1	1/1	1/1	1/1	5
	C+G	0/1	1/1	1/1	1/1	1/1	1/1	5
	C+G+C	0/1	1/1	1/1	1/1	1/1	1/1	5
	C+G+C+C	0/1	1/1	1/1	1/1	1/1	1/1	5
	C+G+C+C+L	0/1	1/1	1/1	1/1	1/1	1/1	5
9	Citrate	1/1	1/1	0/1	1/1	1/1	1/1	5
	Glucose	1/1	1/1	0/1	1/1	1/1	1/1	5
	Cellobiose	1/1	1/1	0/1	1/1	1/1	1/1	5
	Cellulose	1/1	1/1	0/1	1/1	1/1	1/1	5
	Lignin	1/1	1/1	0/1	1/1	1/1	1/1	5
	C+G	1/1	1/1	0/1	1/1	1/1	1/1	5
	C+G+C	1/1	1/1	0/1	1/1	1/1	1/1	5
	C+G+C+C	1/1	1/1	0/1	1/1	1/1	1/1	5
	C+G+C+C+L	1/1	1/1	0/1	1/1	1/1	1/1	5
12	Citrate	1/1	1/1	1/1	1/1	1/1	1/1	6
	Glucose	1/1	1/1	1/1	1/1	1/1	1/1	6
	Cellobiose	1/1	1/1	1/1	1/1	1/1	1/1	6
	Cellulose	1/1	1/1	1/1	1/1	1/1	1/1	6
	Lignin	1/1	1/1	1/1	1/1	1/1	1/1	6
	C+G	1/1	1/1	1/1	1/1	1/1	1/1	6
	C+G+C	1/1	1/1	1/1	1/1	1/1	1/1	6
	C+G+C+C	1/1	1/1	1/1	1/1	1/1	1/1	6
	C+G+C+C+L	1/1	1/1	1/1	1/1	1/1	1/1	6
33	Citrate	2/3	3/3	3/3	3/3	3/3	3/3	17
	Glucose	2/3	3/3	3/3	3/3	3/3	3/3	17
	Cellobiose	2/3	3/3	3/3	3/3	3/3	3/3	17
	Cellulose	2/3	3/3	3/3	3/3	3/3	2/3	16
	Lignin	2/3	3/3	3/3	3/3	3/3	3/3	17
	C+G	2/3	3/3	3/3	3/3	3/3	3/3	17
	C+G+C	2/3	3/3	3/3	3/3	3/3	3/3	17
	C+G+C+C	2/3	3/3	3/3	3/3	3/3	3/3	17
	C+G+C+C+L	2/3	3/3	3/3	3/3	3/3	3/3	17
	Total	37	64	46	64	62	63	336

Extended Data Table 2 |. Consumer-resource model parameters.

Description of each parameter consumer-resource model parameter for all parameterizations. All units are unitless. See Extended Data Fig. 7 for a visual representation of these parameters.

Parameter	Description	Value
i	Consumer (species)
Ni	Abundance of consumer i
α	Resource
Rα	Amount of resource α
Rα0	Resource supply concentration
Dα,β	Resource transformation matrix, indicating which resources β are transformed into α following the consumption.
ci,α	Consumer preference matrix, indicating which resources α are consumed by consumer i.	{0, 1}
l	Leakage fraction, or the fraction of β that becomes α, where the remainder becomes biomass.	0.8
m	Maintenance cost	1
	Number of resources	200 total (T0: 80, T1: 60, T2: 40, T3: 20)
	Number of consumers	1000 total (F0: 500, F1: 300, F2: 100, G: 100)
	Number of resource preferences per consumer	35
	Number of communities	6
	Initial number of consumers per community	200
	Initial abundance of consumers	1
	Number of conditions	7
	Number of resources per condition	20
	Initial abundance of resources	50

Data availability

All raw 16S sequencing data and associated metadata generated for this study can be accessed with the NCBI BioProject accession PRJNA1074799 (https://www.ncbi.nlm.nih.gov/sra/PRJNA1074799). All processed data used in our analyses can be found at https://github.com/segrelab/MetabolicComplexityDivergence.

Code availability

All analysis and simulation code can be found on GitHub at https://github.com/segrelab/MetabolicComplexityDivergence.