Metabolic modeling of microbial communities in the chicken ceca reveals a landscape of competition and co-operation Authors: Irina Utkina, Yi Fan, Benjamin P. Willing, John Parkinson Published: 2025 ABSTRACT BackgroundMembers of the Bacteroidales, particularly Bacteroides species, with their ability to degrade dietary fibers and liberate otherwise unavailable substrates, exert a substantial influence on the microbiome of the lower intestine. However, our understanding of how this influence translates to the metabolic interactions that support community structure remains limited. In this study, we apply constraint-based modeling to investigate metabolic interactions in chicken cecal communities categorized by the presence or absence of Bacteroides.ResultsFrom metagenomic datasets previously generated from 33 chicken ceca, we constructed 237 metagenome-assembled genomes. Metabolic modeling of communities built from these genomes generated profiles of short-chain fatty acids largely consistent with experimental assays and confirmed the role of B. fragilis as a metabolic hub, central to the production of metabolites consumed by other taxa. In its absence, communities undergo significant functional reconfiguration, with metabolic roles typically fulfilled by B. fragilis assumed by multiple taxa. Beyond B. fragilis, we found Escherichia coli and Lactobacillus crispatus also mediate influential metabolic roles, which vary in the presence or absence of B. fragilis. Notably, the microbiome’s compensatory adaptations in the absence of B. fragilis produced metabolic alterations resembling those previously associated with inflammatory bowel disease in humans, including energy deficiency, increased lactate production, and altered amino acid metabolism.ConclusionsThis work demonstrates the potential of using the chicken cecal microbiome as a model system for investigating the complex metabolic interactions and key contributions that drive community dynamics in the gut. Our model-based predictions offer insights into how keystone taxa like B. fragilis may shape the metabolic landscape and functional organization of microbial communities. The observed metabolic adaptations in the absence of B. fragilis share metabolic similarities with profiles seen in dysbiotic states in humans and underscore the translational relevance of these insights for understanding gut health across different host systems. Video AbstractSupplementary InformationThe online version contains supplementary material available at 10.1186/s40168-025-02241-4. Introduction The gut microbiome consists of a complex community of microbes that perform beneficial functions including the extraction of key nutrients (such as short-chain fatty acids (SCFAs), vitamins, and amino acids), promotion of gut health, and limiting pathogen colonization [1, 2]. Helping organize these communities are metabolic interactions involving the sharing (cross-feeding) of metabolites between taxa [3]. Many of these interactions are thought to be mediated by so-called keystone taxa, which play a crucial role in shaping microbial community structure and function [4–9]. Within gut communities, keystone taxa are represented by genera such as Bacteroides and Lactobacillus that have been shown to exert considerable influence within their communities [10, 11]. For example, while Bacteroides spp. play an important role in the degradation of complex carbohydrates and production of SCFAs [12, 13], Lactobacillus spp. have been identified as divisive members, characterized by both positive and negative interactions with other microbes [14]. Currently, our understanding of the cross-feeding relationships that drive community dynamics and the contribution of keystone taxa is limited. Community metabolic modeling offers a powerful approach to unravel the complex interactions within the gut microbiome. Integration of genome-scale metabolic models from multiple microbes within a community-scale framework provides a mechanistic basis to dissect the metabolic landscape of microbial communities and the contributions of individual taxa [15–17]. Recent advances in community-scale metabolic modeling enable researchers to predict how bacteria produce key metabolites such as SCFAs that support gut health [18, 19], identify cross-feeding interactions between bacterial species [15–17], and track changes in metabolite concentrations associated with health and disease [20, 21]. Beyond human health, the gut microbiome has attracted considerable interest from the livestock industry. For example, with global bans on the use of antibiotic growth promoters (AGPs) [22], attention is turning to manipulating gut communities in chickens to address issues of food insecurity associated with a surge in enteric infections [22]. However, little is known about the metabolic interactions driving the formation of stable communities in the poultry gut, creating a critical knowledge gap for developing targeted microbiome interventions. In this study, we leverage community metabolic modeling to investigate the influence of keystone taxa on microbial communities associated with the chicken cecum, a major site of digestion of insoluble fibers and fermentation, and host to the most diverse community in the poultry gut. This model system offers the advantages of natural variation in Bacteroides presence, paired metagenomic and metabolomic data, and an optimal level of complexity for studying community principles while offering translational relevance for poultry health. Applying in silico modeling based on previously published metagenomic data, we investigate the impact of Bacteroides fragilis on community and metabolic dynamics. Our models reveal that B. fragilis significantly impacts microbial diversity and the structure of metabolic interactions. In Bacteroides-depleted communities, we observe an increase in the influence of other taxa including Escherichia coli, Anaerostipes butyraticus, and Lactobacillus crispatus, leading to a functional reorganization of the microbiome. Our findings highlight the pivotal role of the unique metabolic capabilities of B. fragilis and the potential implications for host health and productivity in its absence. Results Bacteroides species influence the microbiome composition and functional capacity To examine the metabolic influence of keystone taxa within cecal microbial communities, we divided 33 samples from a recent study of 7-day-old chickens [13] into two groups: one characterized by the presence of Bacteroides species (HB, n = 15), specifically B. fragilis, and the other without Bacteroides species (NB, n = 18) [13]. While the original study classified the latter group as “Low Bacteroides” (LB) group based on a relative abundance threshold of 0.7%, using a threshold of 0.5% relative abundance, we found that no Bacteroides taxa were present in simulations involving this group, which were subsequently designated as “No Bacteroides” (i.e., NB). Metagenomic datasets previously generated from these samples were used to generate metagenome-assembled genomes (MAGs). To enhance MAG quality and consistency across samples, we conducted both individual assemblies for each sample and a co-assembly of all samples. This combined approach allowed us to capture species present in individual samples and improve assembly quality for shared taxa by leveraging the combined data. Additionally, we re-assembled common taxa by pooling the corresponding reads from multiple samples (see the “Methods” section). This process yielded 237 high- and medium-quality MAGs (defined as having ≥ 50% completeness and ≤ 10% contamination) representing 227 taxa. Of these MAGs, 71 were found exclusively in HB samples, 81 exclusively in NB samples, and 85 were shared between both community types. Each MAG represented at least 0.5% relative abundance in at least one of the 33 samples (Fig. 1A). Communities were largely dominated by representatives of families such as Lachnospiraceae, UBA660, and taxa belonging to the Oscillospirales order, with significant variations in their relative abundances across samples (Fig. S1). Next, we generated genome-scale metabolic models for each MAG, refining models through gap-filling and manual curation to ensure their ability to support growth on a corn feed diet and incorporate missing carbohydrate-active enzymes and fermentation pathways (see the “Methods” section). While these MAGs met our minimum quality thresholds, incomplete genome recovery may underestimate certain functional capabilities, particularly for complex pathways with genes that may have been challenging to assemble accurately from short-read data. Subsequently, models were combined to reflect the cecal samples from which they were derived, resulting in communities represented by 5 to 33 taxa with an average of 1096 ± 98 unique enzymes (as defined by enzyme commission numbers; ECs; Fig. 1A). Further, each community contained from 2187 to 3368 (mean 3001 ± 280) unique reactions encompassing from 143 to 598 (mean 495 ± 83) unique pathways (defined in MetaCyc [23]), with the number of unique pathways being significantly higher in NB communities compared with HB (p = 0.0045, Wilcoxon rank-sum test; Fig. 1B). Our reconstructed metabolic models revealed distinct metabolic profiles among taxa. E. coli had the largest number of enzymes, with broad representation across all pathways, suggesting high functional diversity (Fig. S2; Table S1), although we acknowledge this may also reflect its well-annotated genome. Bacteroides, Alistipes, Helicobacter, Barnesiella, and Coprobacter spp., detected above our abundance threshold (≥ 0.5%), were exclusive to HB communities and exhibited expanded functionality in glycan and vitamin metabolism, suggesting they may play a dominant role in the breakdown of dietary fibers and vitamin synthesis. While most taxa possessed core functionalities across the major pathways, species within UBA1242 (e.g., Caccovivens spp.), UBA660 (e.g., Faecenecus and Faecimonas spp.), and CAG-508 (e.g., Merdicola spp.) families displayed sparser coverage of functions within carbohydrate metabolism. These microbial groups have been recently identified in diverse animal and human gut environments and are characterized by limited metabolic capabilities [24–29]. Despite their distant phylogenetic relationships, they each exhibit reduced genome sizes, diminished genetic repertoires, and numerous auxotrophies [30, 31]. Further, these taxa exhibited minimal overlap with other species in amino acid metabolism capabilities, with distinct sets of enzymes involved in amino acid metabolism and xenobiotic biodegradation, suggesting they occupy a unique niche (e.g., mitigating environmental stressors) within the gut microbiome (Fig. S2).Fig. 1Taxonomic composition and functional analysis of chicken cecal microbial communities. A Relative abundance of MAGs consolidated by taxonomic family across 33 chicken cecal samples. The bar plots represent the numbers of unique species, ECs, unique reactions, and unique pathways based on the metabolic model’s content. Samples are categorized into two groups: those with high Bacteroides abundance (HB, gold) and those without Bacteroides (NB, green). B Distribution of numbers of unique metabolic reactions and pathways in NB and HB samples. Significance determined using the Wilcoxon rank-sum test. C, D PCoA of samples based on abundances of pathways (C) or reactions (D) content (i.e., reaction and pathway counts within each sample were summarized by weighing them by the relative abundance of the species contributing to that reaction or pathway), with dissimilarity matrices calculated using the Bray–Curtis index. Samples are represented by triangles for HB (high Bacteroides, gold) and circles for NB (no Bacteroides, green). Percentages in axis labels indicate the proportion of variance explained by each principal coordinate. E, F PCoA of samples based on the presence or absence of pathways (E) or reactions (F) using Jaccard distances Taxonomic composition and functional analysis of chicken cecal microbial communities. A Relative abundance of MAGs consolidated by taxonomic family across 33 chicken cecal samples. The bar plots represent the numbers of unique species, ECs, unique reactions, and unique pathways based on the metabolic model’s content. Samples are categorized into two groups: those with high Bacteroides abundance (HB, gold) and those without Bacteroides (NB, green). B Distribution of numbers of unique metabolic reactions and pathways in NB and HB samples. Significance determined using the Wilcoxon rank-sum test. C, D PCoA of samples based on abundances of pathways (C) or reactions (D) content (i.e., reaction and pathway counts within each sample were summarized by weighing them by the relative abundance of the species contributing to that reaction or pathway), with dissimilarity matrices calculated using the Bray–Curtis index. Samples are represented by triangles for HB (high Bacteroides, gold) and circles for NB (no Bacteroides, green). Percentages in axis labels indicate the proportion of variance explained by each principal coordinate. E, F PCoA of samples based on the presence or absence of pathways (E) or reactions (F) using Jaccard distances Several taxa exhibited differential abundance patterns between community types. In NB samples, Lactobacillus crispatus (average relative abundance 10 ± 14%), Anaerostipes butyraticus (6.7 ± 8.1%), and E. coli (7.5 ± 5.6%) were more prevalent compared with HB communities (2.7 ± 2.6%, 1%, and 2.8 ± 2.5%, respectively), suggesting these taxa become more dominant in the absence of B. fragilis. Conversely, members of the UBA660 family, specifically Faecenecus gallistercoris and Faecimonas intestinavium, exhibited high relative abundance (15–68%) primarily in NB samples where L. crispatus, A. butyraticus, and E. coli were absent or present at low levels (Fig. S1). These preliminary observations point to potential competitive relationships that we investigate in subsequent analyses. These distribution patterns could indicate scenarios where less metabolically versatile taxa exploit specific substrates or metabolic niches that are less accessible to other bacteria. Members of the order, Oscillospirales, exhibited enhanced functionality in the metabolism of cofactors and vitamins, as well as carbohydrate metabolism, aligning with their roles in fiber degradation and SCFA production (Fig. S2). These diverse metabolic capacities and interdependent roles of bacteria in the chicken cecal microbiome illustrate the complex relationships driving microbial community structure. Principal Coordinate Analysis revealed that the presence of B. fragilis significantly influences both the taxonomic composition and metabolic potential of microbial communities. NB and HB samples clearly segregate when analyzed based on pathway content (Fig. 1C, E) and reaction content (Fig. 1D, F), regardless of whether we assessed abundance-weighted data (Bray–Curtis dissimilarity, Fig. 1C, D), or the presence/absence patterns (Jaccard distance, Fig. 1E, F). The high percentage of variance explained by the first two principal coordinates in pathway analyses (70% for Bray–Curtis, Fig. 1C; 67% for Jaccard, Fig. 1E) together with a significant difference in metabolic pathway profiles (PERMANOVA, p = 0.001) highlights the impact of B. fragilis on community metabolism. This functional segregation demonstrates that taxonomic differences translate to substantial metabolic reconfiguration beyond the mere presence/absence of a single taxon. However, differences in within-group variability between the HB and NB groups were not significant (TukeyHSD test, p = 0.13; permutest, p = 0.11; Fig. S3B). Moreover, we observed distinct pathway and reaction abundance profiles between HB and NB groups (PERMANOVA, p = 0.001 for both), with greater variability within the NB communities (TukeyHSD, p = 0.015 and p = 0.006; permutest, p = 0.009 and p = 0.004, for pathways and reactions, respectively; Fig. S3A, B). Overall, these findings show that B. fragilis has a major influence on metabolic interactions within the gut microbiome. While our results highlight clear functional differentiation between groups, future studies could disentangle the specific contributions of Bacteroides versus other co-occurring taxa to these observed differences. Community metabolic modeling accurately predicts the production of acetate and propionate To further dissect the metabolic contributions of B. fragilis within their communities, we constructed in silico models for each of the 33 samples using the BacArena framework [32] in which the relative abundance of each taxon was defined from the metagenomic datasets. Our in silico model was designed to mimic the physicochemical conditions of the chicken cecum, including low oxygen, neutral pH, peristalsis, and release of cecal contents (see the “Methods” section). To assess the accuracy of our in silico ceca models, simulations were performed for 16 h (representing a cycle of feeding), and predicted SCFA concentrations were compared with those previously measured experimentally [13]. We found a significant correlation between the predicted and experimentally derived concentrations for both acetate (Spearman rho = 0.53, p = 0.0015; Fig. 2A) and propionate (Spearman rho = 0.52, p = 0.0021; Fig. 2A). Moreover, the simulations reflected the experimentally measured differences for the NB and HB communities, although we note the model over-predicted the production of propionate by Bacteroides species (Fig. 2B, Fig. S4). The model was less predictive of butyrate production across communities. This discrepancy likely stems from multiple factors: (1) our model does not account for rapid uptake of microbially produced butyrate by host epithelial cells [33, 34], which affects both community types but may differ in rate between HB and NB communities, particularly due to inflammatory differences observed in the original study [13]; (2) our genome-scale models assume that genetic potential translates to metabolic activity, but regulatory constraints and metabolic precursor availability in vivo may limit actual butyrate production; and (3) we observed butyrate-producing organisms (e.g., Butyricicoccus and Faecalibacterium spp.) in NB samples where measured butyrate was nearly zero. While our genome-based modeling approach assumes these organisms would produce butyrate if they possess the necessary genes, in vivo production likely depends on community context and regulatory factors not captured in our model. Previous studies have demonstrated cross-feeding interactions between Bifidobacterium spp. and Faecalibacterium spp., with acetate produced by the former enhancing butyrate production in the latter [35]. In our simulations, acetate produced by Bifidobacterium pullorum was utilized by other taxa, including Butyricicoccus and Faecalibacterium spp. [36–38], to generate butyrate (Fig. S4). These cross-feeding relationships appear more robust in HB communities, potentially explaining the closer agreement between predicted and actual production of butyrate in these samples. This highlights the importance of studying metabolic interdependencies to better understand SCFA production in the gut microbiome. Overall, our models were able to accurately capture general trends in SCFA production, particularly for acetate and propionate. The models correctly predicted the community-level differences in SCFA profiles between HB and NB groups and accurately represented the underlying cross-feeding interactions that drive these patterns. Butyrate predictions showed greater variability, especially in NB communities, likely due to complex regulatory factors not captured in our metabolic models.Fig. 2Comparison of predicted and measured SCFA concentrations in chicken cecal samples. A Spearman rank correlation between in silico predicted and experimentally measured concentrations of acetate, propionate, and butyrate, with correlation coefficients (Spearman rho) and p-values indicated for each SCFA. Lines are fitted linear regressions, and colored areas around them indicate the 95% confidence interval. Each data point represents a sample. B Bar plots comparing the average predicted (darker bars) and measured (lighter bars) SCFA concentrations in samples with no Bacteroides (NB) and high (HB) Bacteroides abundance. The plots indicate the mean concentrations and standard errors for acetate, propionate, and butyrate. Asterisks indicate statistical significance between NB and HB groups for predicted concentrations and between NB and HB groups for measured concentrations (Wilcoxon rank-sum test, adjusted p-values (Benjamini–Hochberg FDR): *p-adjusted < 0.05, **p-adjusted < 0.01, and ***p-adjusted < 0.001 Comparison of predicted and measured SCFA concentrations in chicken cecal samples. A Spearman rank correlation between in silico predicted and experimentally measured concentrations of acetate, propionate, and butyrate, with correlation coefficients (Spearman rho) and p-values indicated for each SCFA. Lines are fitted linear regressions, and colored areas around them indicate the 95% confidence interval. Each data point represents a sample. B Bar plots comparing the average predicted (darker bars) and measured (lighter bars) SCFA concentrations in samples with no Bacteroides (NB) and high (HB) Bacteroides abundance. The plots indicate the mean concentrations and standard errors for acetate, propionate, and butyrate. Asterisks indicate statistical significance between NB and HB groups for predicted concentrations and between NB and HB groups for measured concentrations (Wilcoxon rank-sum test, adjusted p-values (Benjamini–Hochberg FDR): *p-adjusted < 0.05, **p-adjusted < 0.01, and ***p-adjusted < 0.001 E. coli and B. fragilis drive metabolic dynamics in the chicken cecal microbiome Our analysis of metabolic interactions within chicken cecal microbial communities revealed distinct structural differences between HB and NB communities. To further investigate these differences, we used mutual information (MI) scores to construct a network based on metabolic overlap, where metabolic similarity refers to the extent of shared metabolic reactions between communities. Although not statistically significant (p = 0.174, Wilcoxon rank-sum test with Benjamini-Hochberg (BH) correction), NB communities exhibited higher metabolic similarity than HB communities (pairwise MI scores 0.134 ± 0.076 for NB and 0.123 ± 0.068 for HB; Fig. 3A). Since Bacteroides spp. play a fundamental role as primary fermenters and are often considered keystone taxa in their communities [39–42], their potential monopoly of key resources may drive the reduction in community interactions and metabolic diversity. Moreover, the presence of both B. fragilis and E. coli resulted in further compartmentalization of metabolic functions, as evidenced by significantly lower overlap in metabolic functions (MI score 0.128 ± 0.061) compared with NB communities with E. coli (MI score 0.154 ± 0.08; p = 0.044, Wilcoxon rank-sum test with BH correction) and without E. coli (MI score 0.21 ± 0.11; p = 4e–05, Wilcoxon rank-sum test with BH correction) (Fig. 3A). Interestingly, samples lacking E. coli species exhibited the highest pairwise MI scores (0.213 ± 0.111 and 0.21 ± 0.11, for HB and NB communities, respectively), indicating that absence of E. coli promotes higher metabolic similarity irrespective of the presence or absence of B. fragilis. These findings suggest that consistent with its wide range of metabolic functions, E. coli acts as a metabolic generalist—a species capable of performing a broad range of metabolic activities, in contrast to specialists that have more limited, niche-specific roles—to fulfill roles vacated by other keystone taxa to support the metabolic versatility of the cecal microbiome (Fig. S2;Table S1). While MI scores did not show a significant difference in metabolic similarity in NB and HB communities, the influence of additional taxa such as E. coli led us to investigate whether the absence of B. fragilis results in communities exhibiting a greater diversity of less dominant taxa, with increased competition and metabolic interactions. We applied three complementary metrics (see the “Methods” section): Metabolic Resource Overlap (MRO) score, a metric that quantifies both competition for shared resources as well as complementarity in metabolic capabilities [43]; Community Metabolic Dissimilarity (CMD) score, a metric that quantifies the diversity of metabolic functions in a community; and Cross-Feeding Coefficient (CFC) based on the number of metabolite exchanges. We found that NB communities had significantly higher MRO (p = 1.2e–05, Wilcoxon rank-sum test) and CFC (p = 0.025, Wilcoxon rank-sum test) scores than HB communities (Fig. 3B, C). While NB communities also featured higher CMD scores, consistent with reports showing that metabolic dissimilarities enhance cross-feeding between species [44], this difference was not significant. Overall, these findings show that NB communities exhibit greater metabolic similarity between communities and significantly enhanced cross-feeding relationships, supporting the hypothesis that the presence of B. fragilis reduces metabolic interactions and potentially affects metabolic diversity within its community. These differences in interaction patterns between community types should be interpreted as alternative organizational states rather than as indicators of improved community stability or function in the absence of B. fragilis.Fig. 3E. coli and B. fragilis influence metabolic similarity and cross-feeding in chicken cecal microbiomes. A. Mutual information-based network of chicken cecal microbial communities, highlighting metabolic similarities based on metabolic reaction content. Large nodes represent samples, with their sizes corresponding to the number of unique reactions in a sample, colored by category: no Bacteroides (NB, green) and high Bacteroides (HB, gold). Edges are colored and thickened according to the mutual information score, with thicker and darker edges indicating higher similarity. The top region outlined in red indicates samples lacking E. coli. Sizes of the smaller nodes within a sample, representing taxonomic families, correspond to the number of reactions covered by the metabolic models of species belonging to the respective taxonomic families (grouped by taxonomic order below). B. Bar plots showing community sizes (top) and box plots showing Metabolic Resource Overlap (MRO) within each sample (summarizing five in silico replicates). NB communities exhibit significantly higher MRO (p = 1.2e-05, Wilcoxon Rank-Sum test). C. Cross-feeding coefficient (left) and Community Metabolic Dissimilarity (CMD) score (right) between NB and HB communities. NB communities show a significantly higher cross-feeding coefficient (*p=0.025) and a higher average CMD score E. coli and B. fragilis influence metabolic similarity and cross-feeding in chicken cecal microbiomes. A. Mutual information-based network of chicken cecal microbial communities, highlighting metabolic similarities based on metabolic reaction content. Large nodes represent samples, with their sizes corresponding to the number of unique reactions in a sample, colored by category: no Bacteroides (NB, green) and high Bacteroides (HB, gold). Edges are colored and thickened according to the mutual information score, with thicker and darker edges indicating higher similarity. The top region outlined in red indicates samples lacking E. coli. Sizes of the smaller nodes within a sample, representing taxonomic families, correspond to the number of reactions covered by the metabolic models of species belonging to the respective taxonomic families (grouped by taxonomic order below). B. Bar plots showing community sizes (top) and box plots showing Metabolic Resource Overlap (MRO) within each sample (summarizing five in silico replicates). NB communities exhibit significantly higher MRO (p = 1.2e-05, Wilcoxon Rank-Sum test). C. Cross-feeding coefficient (left) and Community Metabolic Dissimilarity (CMD) score (right) between NB and HB communities. NB communities show a significantly higher cross-feeding coefficient (*p=0.025) and a higher average CMD score Functional reorganization and emergence of alternate hubs in response to the absence of keystone taxa in the gut Building upon the finding that the presence or absence of B. fragilis influences cross-feeding relationships among other taxa, we extended our analyses to examine changes in metabolite exchange dynamics in a larger set of taxa that could be considered metabolic hubs (defined here as taxa exchanging at least 12 metabolites with at least 10 other taxa across samples within each category). We identified 87 such metabolic hub taxa in total, with 40 associated exclusively with NB communities, 22 with HB communities, and 25 present in both community types. These taxa were used to generate profiles of interactions with other taxa and subsequently clustered to identify cliques of cross-feeding taxa (Fig. 4A; Table S2). Consistent with our previous findings B. fragilis represents the central hub in the HB communities, while in its absence, E. coli, A. butyraticus, and L. crispatus adopt a greater range of metabolic interactions. The shift toward a more partitioned network observed in NB communities was further characterized by the formation of distinct cliques, representing groups of taxa engaging in mutual cross-feeding interactions. Within each clique, a single taxon (e.g., Mediterraneibacter sp904377845, Lactobacillus johnsonii, and Borkfalkia sp944368795) appears to dominate interactions acting as a metabolic hub for each group of taxa. To further clarify how “metabolic hub” taxa exert their influence within their communities, we examined their interactions with other taxa in the context of prevalence across samples, relative abundance, and provision of unique functionality (Fig. 4B, Fig. S5). Surprisingly, a taxon’s unique metabolic potential within its community (as estimated by the number of unique enzymes) did not correlate with its number of metabolic interactions (Fig. S5A) although we did observe a positive correlation with prevalence (Fig. S5B). These findings suggest that while the number of unique functions a taxon brings to a community might increase its ability to engraft, the community does not exploit that additional functionality. It is important to note that the simple prevalence of a taxon across samples indicates successful exploitation of its metabolic potential for its own benefit. However, our analysis distinguishes between this individual success and how a taxon’s metabolic capabilities benefit the broader community through cross-feeding relationships—two complementary perspectives that together provide a more complete understanding of how this ecosystem functions. Indeed, E. coli, which is both prevalent and possesses the greatest set of unique functions, appears to benefit in its community from the consumption of many metabolites from multiple species with only modest contributions to production (Fig. 4B). Other prevalent taxa exhibiting a broad range of interactions, such as L. crispatus and A. butyraticus, with fewer unique enzymes did not exhibit the same disparity in consumption and production. At the other end of the spectrum, we found some less prevalent taxa, present in only a limited number of samples (e.g., L. johnsonii and M. sp904377845) exert their influence by specializing in the production of a wide array of metabolites utilized by a large number of other taxa (Fig. 4B). While E. coli’s extensive functional annotation likely contribute to its apparent metabolic versatility, its designation as a metabolic hub in our analysis is based primarily on observed interaction patterns and ecological relationships rather than gene content alone.Fig. 4Major metabolic hubs and cross-feeding dynamics in HB and NB microbial communities. A Heatmap of interactions between species (clustered hierarchically using complete linkage and euclidean distance, with rows split by HB/NB presence pattern and columns split by sample type (HB/NB), with clustering performed within each split), highlighting top cross-feeders defined as those forming at least 10 cross-feeding pairs with the exchange of a minimum of 12 metabolites across samples within each category, HB or NB. The interactions observed across multiple samples within each category were normalized by the number of samples in which each cross-feeding pair was present. Key taxa are labeled in this heatmap, including taxa present in at least two samples, forming an average of at least 230 interactions per sample and possessing, on average, at least 5 unique EC per sample. The tiles in the heatmap are colored based on the normalized total number of interactions for each species pair, with a color gradient indicating the interaction intensity. Blocks of taxon interactions represent“cliques,”indicating groups of species with frequent metabolic exchanges. For example, the green-highlighted clique shows Lactobacillus johnsonii acting as a “metabolic hub,” engaging in extensive interactions with taxa from the families Lachnospiraceae (Lachnoclostridium B sp000765215, Fusicatenibacter sp900543115, Blautia ornithocaccae, Eisenbergiella sp904392525, Choladocola avistercoris, Merdisoma faecalis), Acutalibacteriaceae (Eubacterium R faecavium, Acutalibacter ornithocaccae, Anaeromassilibacillus stercoravium), Borkfalkiaceae (Coproplasma stercoravium, Coproplasma stercorigallinarum), Anaerotignaceae (Anaerotignum merdipullorum), Butyricicoccaceae (Butyricicoccus avicola), Ruminococcaceae (Gemmiger avicola), and Oscillospiraceae (Dysosmobacter pullicola), as well as with each other. Taxa mentioned in the main text are highlighted in bold. B Stacked bar charts representing the intensity of cross-feeding per taxon. The bars indicate the average number of interacting species for each taxon per sample, divided into metabolite consumption (left) and production (right). Colors correspond to the number of metabolites exchanged, with categories ranging from 1–5, 6–10, 11–15, and > 15 metabolites. The bar plots to the right represent the sample count (number of samples in which each species is present, blue), relative abundance (gray), and the average number of unique enzymes (ECs) a taxon possesses per sample (purple). Subset of taxa displayed is present in at least two samples, exhibits an average of at least 230 total interactions per sample, and possesses a minimum of five unique EC numbers per sample Major metabolic hubs and cross-feeding dynamics in HB and NB microbial communities. A Heatmap of interactions between species (clustered hierarchically using complete linkage and euclidean distance, with rows split by HB/NB presence pattern and columns split by sample type (HB/NB), with clustering performed within each split), highlighting top cross-feeders defined as those forming at least 10 cross-feeding pairs with the exchange of a minimum of 12 metabolites across samples within each category, HB or NB. The interactions observed across multiple samples within each category were normalized by the number of samples in which each cross-feeding pair was present. Key taxa are labeled in this heatmap, including taxa present in at least two samples, forming an average of at least 230 interactions per sample and possessing, on average, at least 5 unique EC per sample. The tiles in the heatmap are colored based on the normalized total number of interactions for each species pair, with a color gradient indicating the interaction intensity. Blocks of taxon interactions represent“cliques,”indicating groups of species with frequent metabolic exchanges. For example, the green-highlighted clique shows Lactobacillus johnsonii acting as a “metabolic hub,” engaging in extensive interactions with taxa from the families Lachnospiraceae (Lachnoclostridium B sp000765215, Fusicatenibacter sp900543115, Blautia ornithocaccae, Eisenbergiella sp904392525, Choladocola avistercoris, Merdisoma faecalis), Acutalibacteriaceae (Eubacterium R faecavium, Acutalibacter ornithocaccae, Anaeromassilibacillus stercoravium), Borkfalkiaceae (Coproplasma stercoravium, Coproplasma stercorigallinarum), Anaerotignaceae (Anaerotignum merdipullorum), Butyricicoccaceae (Butyricicoccus avicola), Ruminococcaceae (Gemmiger avicola), and Oscillospiraceae (Dysosmobacter pullicola), as well as with each other. Taxa mentioned in the main text are highlighted in bold. B Stacked bar charts representing the intensity of cross-feeding per taxon. The bars indicate the average number of interacting species for each taxon per sample, divided into metabolite consumption (left) and production (right). Colors correspond to the number of metabolites exchanged, with categories ranging from 1–5, 6–10, 11–15, and > 15 metabolites. The bar plots to the right represent the sample count (number of samples in which each species is present, blue), relative abundance (gray), and the average number of unique enzymes (ECs) a taxon possesses per sample (purple). Subset of taxa displayed is present in at least two samples, exhibits an average of at least 230 total interactions per sample, and possesses a minimum of five unique EC numbers per sample These findings reveal a hierarchy of taxa acting as metabolic hubs, where taxa such as B. fragilis and E. coli are responsible for driving the bulk of interactions, supplemented by additional hubs (e.g., A. butyraticus and L. crispatus). Further, within their communities, there exist cliques of taxa sharing metabolic interactions, organized around relatively rare taxa (present in ≤ 5 samples, compared with 10–20 samples for more common taxa). This hierarchy highlights the varying abilities of each type of hub taxa to colonize within their communities, with B. fragilis and E. coli capable of colonizing many communities and establishing extensive cross-feeding networks. Others, such as M. sp904377845 and L. johnsonii, appear more limited in their ability to colonize communities and feature a smaller range of interactions with only a subset of taxa. Role of key metabolic hubs in modulating cross-feeding interactions To further assess how the presence/absence of different metabolic hubs may impact cross-feeding relationships within their communities, we constructed a linear model based on the abundance of E. coli and the presence/absence of three additional metabolic hub taxa (A. butyraticus, L. crispatus, and B. fragilis; see the “Methods” section). A higher relative abundance (> 5%) of E. coli abundance was associated with increased cross-feeding in Bacteroides-depleted communities, potentially indicating the ability of E. coli to replace community functions otherwise performed by B. fragilis (Fig. S5C; Table S3). Interestingly, only one of the 15 HB samples exhibited such an abundance, suggesting a negative relationship between these two taxa. Extending this analysis to two other metabolic hub taxa (A. butyraticus and L. crispatus), we found their presence in the absence of E. coli was associated with increased cross-feeding activity, suggesting their ability to provide an alternative source of metabolic functions (Fig. S5D-E; Table S3). Communities featuring both E. coli together with A.butyraticus displayed lower cross-feeding activity, suggesting possible overlapping metabolic niches between these taxa. In contrast, NB communities with L. crispatus showed higher cross-feeding activity compared with HB communities, regardless of E. coli abundance, pointing to a potential antagonistic relationship between L. crispatus and Bacteroides. However, it is important to note that while these patterns are statistically supported, they may also be a consequence of other mechanisms, including direct resource competition, indirect interactions mediated through other community members, differential responses to environmental factors, and host-mediated selection processes. Since our linear model focused only on taxa representing metabolic hubs, we were interested in how other metabolic hub taxa, specifically those mediating interactions within a subset of taxa (cliques), contribute to their communities. To avoid the confounding influence of dietary metabolites, we further restricted these analyses to metabolites only produced by bacteria. As before, B. fragilis represents the central hub in the HB communities, while E. coli, A. butyraticus,and L. crispatus assume additional interactions in NB communities (Fig. S6A; Table S4). Moreover, we also observed the formation of metabolic cliques, with greater numbers of interactions in NB communities. Comparing the eight largest cliques specific to either NB or HB communities (labeled A–H), we identified 48 unique microbially derived metabolites exchanged within these cliques (ranging from 24 in clique H to 37 in clique A). While taxa associated with these cliques may not necessarily be present in the same samples, these cliques nonetheless represent groups of taxa that occupy a similar metabolic niche—that is, they share specific sets of metabolic functions and interactions they engage in, contributing to similar ecological roles across different communities. This aligns with previous research demonstrating enrichment of mutualistic interactions in co-occurring microbial communities [43]. While there is a substantial overlap in metabolites exchanged across the eight cliques investigated—17 metabolites were common to all cliques, and an additional 16 were shared among at least five cliques (Fig. S6B)—individual cliques exhibit unique metabolic exchanges reflecting the specific taxa involved. This suggests that, despite a shared core of metabolic pathways and intermediates, each clique possesses distinct metabolic interactions driven by its constituent taxa. For instance, in the NB-specific clique B, Bifidobacterium spp. and members of Lactobacillaceae uniquely produced and exchanged d-alanine, melibiose, and spermidine (Fig. S6A, B). Relative to other cliques, clique D, comprising members of Oscillospirales, Lachnospirales, Christensenellales, Coriobacteriales, and RF39, exhibited a more limited range of metabolite exchanges, lacking common fermentation products but uniquely exchanging NAD, highlighting the role of redox reactions and energy metabolism within this clique. These findings highlight how specific taxa contribute to functional diversity within microbial communities by occupying distinct metabolic niches, thereby shaping the overall metabolic landscape of their environments. The changes in cross-feeding relationships associated with the absence of B. fragilis led us to investigate how this translates in terms of the exchange of specific metabolites. We therefore analyzed the production and consumption of 18 metabolites selected on the basis of their relationship with four keystone taxa (see the “Methods” section; Fig. 5). Our models revealed distinct nutrient extraction patterns between communities. As expected, in HB communities, B. fragilis is the predominant producer, contributing to the majority of the production of 10 metabolites, with E. coli, Mediterraneibacter glycyrrhizinilyticus, Faecalibacterium gallistercoris, and Dysosmobacter pullicola largely responsible for the production of a more limited set of metabolites including methanol, indole, uridine, formate, and l-glutamine. In NB communities, the production of these metabolites is redistributed to multiple other species. For example, consistent with previous reports [35, 38, 45, 46], in the absence of B. fragilis, E. coli assumes the production of multiple metabolites otherwise produced by B. fragilis including acetate (along with members of Clostridia), ammonium (with A. butyraticus), succinate, propionate, sorbitol, and GABA. These activities are supported through the production of lactate, uridine, and l-glutamine by L. crispatus, the latter consistent with previous reports of its production [47]. In the presence of B. fragilis, E. coli alters its metabolic contributions to the generation of methanol, itself utilized by B. fragilis. Interestingly, in NB communities, A. butyraticus adopts the role of methanol production (Fig. 5). While our metabolite exchange analyses provide evidence that E. coli assumes certain metabolic functions in the absence of B. fragilis, future work directly comparing their metabolic route selections could help distinguish whether E. coli functionally replaces B. fragilis through similar enzymatic pathways or opportunistically expands into vacated niches through alternative metabolic strategies. Together, our predictions show that in the absence of a keystone taxon (here, B. fragilis), the cecal microbial community undergoes significant reorganization, characterized by the emergence of alternate metabolic hubs as well as a shift in metabolite production and consumption profiles by taxa present in both types of communities. This redistribution of nutrient extraction capabilities affects the overall metabolic efficiency and may have downstream consequences for host nutrition and health.Fig. 5Shifts in cross-feeding networks: metabolic role redistribution in HB and NB communities. The Sankey diagram illustrates the cross-species interactions between primary producers and consumers of key cross-fed metabolites in the HB and NB communities. Producers are displayed on the left, cross-fed metabolites in the middle, and consumers on the right. Gold links represent contributions in HB communities, whereas green links represent those in NB communities. The thickness of each link is proportional to the contribution of the species (as a percentage) to the total production or consumption of each metabolite. To highlight the major players in cross-feeding dynamics, only the top three producers and consumers for each metabolite are shown, with a cutoff set at a minimum 5% contribution across samples within each category. Species highlighted in bold were previously described in Fig. 4B. The diagram provides a visual summary of how key microbial taxa participate in the metabolic exchange, highlighting functional shifts between HB and NB communities Shifts in cross-feeding networks: metabolic role redistribution in HB and NB communities. The Sankey diagram illustrates the cross-species interactions between primary producers and consumers of key cross-fed metabolites in the HB and NB communities. Producers are displayed on the left, cross-fed metabolites in the middle, and consumers on the right. Gold links represent contributions in HB communities, whereas green links represent those in NB communities. The thickness of each link is proportional to the contribution of the species (as a percentage) to the total production or consumption of each metabolite. To highlight the major players in cross-feeding dynamics, only the top three producers and consumers for each metabolite are shown, with a cutoff set at a minimum 5% contribution across samples within each category. Species highlighted in bold were previously described in Fig. 4B. The diagram provides a visual summary of how key microbial taxa participate in the metabolic exchange, highlighting functional shifts between HB and NB communities Metabolic reorganization associated with the loss of Bacteroides in the chicken cecum parallels metabolic features observed in dysbiotic states Our comparisons of HB and NB communities revealed metabolic shifts that share features with altered microbial metabolism observed in dysbiotic states, including those reported in human inflammatory bowel disease (IBD, which includes ulcerative colitis—UC and Crohn’s disease—CD). While acknowledging the significant differences in gut structure, immune function, and microbial composition between avian and mammalian species, we note that these metabolic reorganization patterns may reflect conserved community responses to keystone taxa depletion across different biological systems. Previous studies have reported shifts in the abundance and roles of both B. fragilis and E. coli—major metabolic hubs in our cecal communities—that correlate with disease states. The former is associated with reduced abundance in IBD patients, whereas the latter is associated with increased abundance [48–52]. Additionally, we observed increased abundance of Alistipes and Faecalibacterium species in HB samples (Fig. S1), consistent with communities associated with healthy individuals in contrast to IBD patients [51, 53, 54]. Some of the metabolic changes we observe between HB and NB communities parallel certain metabolic patterns that have been reported in studies comparing healthy individuals and IBD patients (albeit, as noted above, with important caveats regarding cross-species comparisons). For example, the heterolactic fermentation pathway associated with lactic acid bacteria is more prevalent in NB communities, resulting in increased starch consumption and lactate production (Fig. 6), consistent with elevated lactate levels observed in IBD patients [55–57]. NB communities are also associated with reduced production of propionate, a key source of energy for gut epithelial cells and a hallmark of IBD [57–60]. While studies have reported varied metabolic profiles in IBD, with both increased and decreased energy metabolism [57, 61–63], our findings align with the broader trend of altered energy metabolism in the disease state. Specifically, we observed a metabolic shift in Bacteroides-depleted communities characterized by increased consumption of coenzyme A (CoA) and glycerol, alongside increased production of TCA cycle intermediates such as 2-oxoglutarate, as well as GTP. CoA, essential for fatty acid metabolism and the TCA cycle as an acyl group carrier, is being consumed more by the microbial community, suggesting upregulated lipid and energy metabolism pathways. However, the concurrent accumulation of TCA intermediates may indicate a bottleneck or dysregulation within the microbial TCA cycle, potentially due to limited CoA availability. This imbalance could reflect a shift toward fermentative metabolism in the absence of B. fragilis. Notably, reduced CoA levels have been observed in the colonic tissues of pig models and human patients with ulcerative colitis (UC), where they are associated with impaired fatty acid oxidation and energy metabolism in host cells [57, 64]. While our findings pertain to increased microbial consumption of CoA, this could potentially decrease its availability in the gut lumen, impacting the host’s energy metabolism.Fig. 6Comparison of metabolite exchange patterns between HB and NB communities. Each dot represents a metabolite being exchanged, with black dots indicating metabolites that do not show statistically significant changes in production or consumption between NB and HB groups. Statistically significant differences in total community fluxes (normalized to community size) were identified using the Wilcoxon rank-sum test. Dashed red line indicates the threshold for significance level, FDR-corrected p-value < 0.05. Metabolites are color-coded based on the direction of their exchange: green - increased consumption in HB; blue - increased production in HB; orange - increased consumption in NB; purple - increased production in NB; red - switch from production in NB to consumption in HB; pink - switch from consumption in NB to production in HB. Metabolites with an absolute log2FoldChange greater than 0.7 are labeled and colored accordingly; metabolites discussed in results are highlighted in bold Comparison of metabolite exchange patterns between HB and NB communities. Each dot represents a metabolite being exchanged, with black dots indicating metabolites that do not show statistically significant changes in production or consumption between NB and HB groups. Statistically significant differences in total community fluxes (normalized to community size) were identified using the Wilcoxon rank-sum test. Dashed red line indicates the threshold for significance level, FDR-corrected p-value < 0.05. Metabolites are color-coded based on the direction of their exchange: green - increased consumption in HB; blue - increased production in HB; orange - increased consumption in NB; purple - increased production in NB; red - switch from production in NB to consumption in HB; pink - switch from consumption in NB to production in HB. Metabolites with an absolute log2FoldChange greater than 0.7 are labeled and colored accordingly; metabolites discussed in results are highlighted in bold Beyond changes in energy metabolism, simulations also predict an increase in production of amino acids, including l-phenylalanine, l-valine, and l-lysine in NB communities relative to HB (Fig. 6), consistent with recent studies reporting alterations in levels of these amino acids in IBD patients [55, 59, 65, 66]. Other notable metabolic similarities include reduced profiles of 5-methylthio-d-ribose, a metabolite involved in the methionine salvage pathway, in NB communities, indicative of reduced sulfur cycle activity and consistent with deficiencies in sulfur-containing metabolites in microbial communities of IBD patients [20, 21]. Thiamin, typically deficient in CD patients [55, 67], is predominantly produced by B. fragilis, and consequently reduced in NB communities. Finally, the increased consumption of sialic acid (Neu5Ac), a mucin-derived monosaccharide, by NB microbiomes suggests enhanced sialidase activity, which has been linked to intestinal inflammation and provides a growth advantage to bacteria like E. coli [68, 69]. Thus, while the earlier study reported reduced inflammation in HB chicken cecal microbiomes [13], the opposite may be inferred for NB microbiomes. Overall, the metabolic reorganization observed in our chicken cecal NB communities reflects aspects of metabolic features associated with dysbiotic states of human gut communities. While it is important to recognize that these comparisons are limited to metabolic patterns and do not imply disease equivalence between avian and mammalian systems, they nevertheless reveal that the loss of keystone taxa like B. fragilis may trigger similar metabolic adaptations across different host systems. Discussion Previous studies suggest microbial communities in the gut display a hierarchical structure, with so-called keystone taxa acting as influential members through their ability to provide their communities with unique metabolic capacities [9, 70–72]. The Bacteroidaceae family members are considered to be one of the most influential taxa, exerting a significant impact on metabolic interactions within their communities [39–41, 73, 74]. Previous studies have shown that the poultry gut, specifically the cecum, exhibits substantial fluctuations in the abundance of members of the Bacteroidaceae family, including the genus Bacteroides [13, 75, 76]. Notably, Bacteroides spp. are less abundant in intensively raised chickens compared to those raised in extensive systems (i.e., with access to the outdoors) [77]. Shifts in the abundance of Bacteroides can impact the community dynamics resulting in reduced feed efficiency and heightened susceptibility to enteric diseases [78–80]. In humans, such shifts have also been associated with IBD [81, 82]. Here, we leveraged a pre-existing dataset of metagenomic data from the cecal communities of chickens displaying a range of Bacteroides abundance and performed community metabolic modelling to dissect the influence of keystone taxa on metabolic interactions within their communities. Notably, we found that our sample-specific community models accurately predicted the production of SCFAs, particularly acetate and propionate, and recapitulated experimentally observed differences between Bacteroides-rich (HB) and -depleted communities (NB). Our findings indicate that communities lacking B. fragilis exhibit more similar metabolic profiles across different microbiomes while simultaneously displaying a wider range of metabolic functions within each individual community. However, this shift does not necessarily translate to an increase in overall microbial diversity. Moreover, NB communities exhibited significantly higher metabolic resource overlap within communities and increased cross-feeding interactions. These observations are consistent with B. fragilis functioning as a keystone taxon that dominates and monopolizes key metabolic functions [83, 84]. In the absence of B. fragilis, these functions appear to be distributed across multiple organisms, necessitating more extensive metabolic exchanges. By defining metabolic hubs as microbial taxa with a central role in metabolite exchange and distribution, we examined their influence on the overall metabolic landscape. While B. fragilis acts as a central hub, its absence triggers a shift toward a decentralized network characterized by distinct cross-feeding cliques and the emergence of new metabolic hubs, such as E. coli, A. butyraticus, and L. crispatus. This reorganization represents different metabolic architectures rather than indicating improved or diminished community function. While centralized metabolism around B. fragilis and distributed metabolism across multiple taxa in NB communities might each offer distinct functional characteristics, our current data describes these organizational differences without supporting conclusions about their relative effects on community stability or host health. Contrary to the expectation that a larger enzymatic repertoire directly translates to greater community influence, we observed a diverse range of strategies, from generalists like E. coli, which can perform a broad range of metabolic activities, to specialists like L. johnsonii, which focus on specific metabolic functions. Our analyses show that E. coli has a significant impact on cross-feeding, with increased activity when E. coli abundance exceeds 5% in Bacteroides-depleted communities. Interestingly, L. johnsonii and L. crispatus, previously described as eliciting a Marmite effect on the gut microbiome [14], emerged as alternative metabolic hubs in the absence of Bacteroides, reflecting the antagonistic relationship between Lactobacilli and Bacteroides that has been observed in previous studies [13, 85–87]. While L. crispatus had a positive effect on cross-feeding in the absence of E. coli, its presence in samples where E. coli was abundant led to a decrease in metabolic interactions. These observations highlight the shifting role of L. crispatus as it can both contribute to and limit the community’s metabolic interactions, depending on the presence of keystone taxa (e.g., E. coli and B. fragilis). In our study, we found that B. fragilis plays a central role in the production of metabolites utilized by other taxa, whereas NB communities were characterized by a greater distribution of metabolic functions across a broader range of taxa. The metabolic reorganization we observed in Bacteroides-depleted chicken communities shares certain features with altered microbial metabolism reported in human IBD, such as increased lactate production, altered energy metabolism, and changes in amino acid profiles. However, these similarities should be interpreted cautiously, given the differences in gut physiology and immune function between avian and mammalian species. We acknowledge that the literature on Bacteroides in IBD is inconsistent, with varying reports of depletion or enrichment across different studies and disease subtypes. Recent work has revealed complex relationships between B. fragilis and IBD, with enterotoxigenic strains often associated with disease exacerbation, whereas non-toxigenic strains may provide metabolic benefits that protect against inflammation [88]. While chickens should not be considered direct models for human disease, as we have shown here, studying metabolic reorganization across diverse host systems can deepen our understanding of how microbial communities function with or without specific keystone taxa. These patterns of adaptation may represent fundamental principles of microbial community organization in different ecological contexts. This work establishes a framework for utilizing community metabolic modeling to study the poultry gut microbiome. Our approach facilitates the prediction of metabolic outputs, such as SCFA production, as well as elucidation of cross-feeding relationships. Moreover, our framework offers a mechanistic understanding of how keystone taxa, like Bacteroides, influence the ecological dynamics and metabolic landscape of the gut microbiome. Despite its promising results, we acknowledge our platform is not without some limitations. A specific limitation of our approach is the potential overattribution of metabolic activities to well-characterized taxa, such as E. coli. This is a common challenge in metabolic modeling, where species with more complete genome annotations and metabolic reconstructions may show exaggerated functional capabilities [89, 90]. Nevertheless, our core findings remain robust for several reasons. First, our community-level predictions of SCFA concentrations align with experimental measurements, validating the biological relevance of our models despite potential individual species biases. Second, the consistent metabolic patterns observed across multiple independent samples demonstrate reproducible community behaviors. Third, our analysis of metabolic reorganization at the community level captures fundamental functional shifts beyond individual species contributions. Finally, the predicted roles of B. fragilis and E. coli as metabolic hubs are supported by their documented metabolic versatility and capacity to process diverse substrates [41, 42, 91, 92]. These factors collectively demonstrate that while individual contributions may be imperfectly estimated, the broader community-level dynamics and functional shifts we describe represent genuine biological phenomena. Additionally, future studies could further quantify the metabolic interactions observed in our system using complementary approaches such as the interaction scoring system recently developed by Lambert et al. [93], which could provide additional insights into the mechanisms driving the metabolic relationships identified in this study. An additional caveat in this study concerns the absence of host-microbiota interactions, which ignores the potential contribution of the host in terms of uptake of metabolites, as well as the influence of the host immune system. While the models accurately predict overall trends in metabolite production, they may overestimate or underestimate specific metabolic fluxes. This can be attributed to limitations in the functionality of some bacterial models, influenced in turn by the quality of the corresponding MAGs. Varying MAG completeness may lead to underestimation of certain metabolic capabilities, particularly for pathways where key genes fall in regions difficult to assemble. This potentially affects our metabolic similarity analyses, as less complete MAGs will result in fewer shared functions with other community members. Moreover, MAG-based models have two inherent limitations. First, they may not include mobile genetic elements that often encode metabolically significant functions, potentially underestimating the community’s metabolic capabilities. Second, they represent genomic potential rather than actual metabolic activity; this impacts the ability to assign reaction flux constraints during modeling that can lead to an inaccurate representation of pathway activities. Finally, the lack of data from upstream gut compartments may also limit our understanding of the precise metabolic environment in the cecum; here, we assume that dietary components in amounts adjusted for known estimates of upstream absorption are directly available to members of the cecal community, ignoring the potential metabolic contributions from microbes present upstream of the ceca. Future research could integrate host factors, explore other gut regions, and assess the potential for targeted interventions to beneficially manipulate the microbiome. With standardized diets and housing regimens, the chicken cecal model used here provides a relatively simple system to examine the complex metabolic interactions driving microbial community dynamics. However, we believe our framework could be readily adapted for studies focused on human gut communities by accounting for the greater complexity of the human gut. Specific challenges include less characterized regional absorption profiles, higher inter-individual variability requiring more extensive parameterization, and the need for more assumptions regarding metabolite availability along the extended digestive tract. Methods External data collection Publicly available raw shotgun metagenomic sequence data from 33 chicken ceca and metadata associated with the study [13] were downloaded from the sequence read archive (SRA; BioProject ID PRJNA902117). Chicken cecal metagenomic samples analysis Summary statistics of raw metagenomic sequencing data for the 33 samples included in the study, including total number of reads, total number of bases (dataset size), and read length distributions, are provided in Table S6. Metagenomic reads were subjected to quality filtering using BBTools/BBMap (v39.03) [94]. Single-sample assemblies and co-assemblies of short reads were conducted with MEGAHIT (v1.2.9) [95], specifying a minimum contig length parameter of “–min-contig-len” set to 1000. The Bowtie2 (v2.2.3) [96] tool was employed to index assemblies and map filtered reads onto the contigs, creating abundance profiles. Following this, the metagenome assemblies underwent separate binning processes utilizing MetaBAT2 (v2.13) [97] and CONCOCT (v1.1.0) [98]. Initially, samtools (v1.9) [99] sorted the read mappings, and the read coverage was determined using the MetaBAT2 script “jgi_summarize_bam_contig_depths.” Subsequently, the binning process involved the “metabat2” command with default settings for each assembly, using contig coverage across samples as input. For CONCOCT, accessory scripts were employed to cut contigs into 10 kb fragments (cut_up_fasta.py), and the read coverage for these fragments was computed using CONCOCT_coverage_table.py. The fragment coverages were then utilized for binning the 10 kb fragments, and the clustered fragments underwent merging (merge_cutup_clustering.py) to generate the final CONCOCT bins (extra_fasta_bins.py). To dereplicate and consolidate the produced outputs from CONCOCT and MetaBAT2 binners, the metaWRAP (v1.2.3) [100] “bin refinement” command was employed, utilizing parameters “-x 10 -c 50.” These parameters signify a maximum bin contamination threshold of 10% and a minimum bin completeness threshold of 50%, as determined by CheckM [101] estimates. Taxonomic annotation of generated bins was performed using GTDB-Tk (v2.3.2) [102] command “gtdbtk classify_wf” with default settings. Finally, the abundances of the draft genomes (bins) across the samples were quantified using the metaWRAP “quant_bins” module. To improve the consistency and quality of models representing the same species across different samples, different assembly strategies were employed depending on species prevalence and strain heterogeneity. Since all samples were derived from chickens raised in the same environment, we expected limited strain variation for most species, making it reasonable to pool reads from different samples. For species present in multiple samples (except E. coli), reads corresponding to each species were pooled from all samples where the species was detected and reassembled, followed by re-binning using the previously described procedure to generate a single MAG per species (Fig. S3D; Table S5). E. coli required a different approach due to its high prevalence (present in 23 samples) and strain heterogeneity. Initially, pooling all E. coli reads resulted in an 8.2 Mb assembly, well above the typical E. coli genome size, suggesting the presence of multiple strains. To represent the multiple E. coli strains present across samples without retaining the 23 sample-specific E. coli individual assemblies with varying size and completeness, the bins derived from the individual assemblies were clustered based on average nucleotide identity (ANI) using the fastANI software [103]. A distance matrix was constructed, and the bins were then clustered using the hclust function from the R-package stats. Five clusters were selected, and five cluster-specific E. coli strains were re-assembled, each representing a species pooled from samples within the corresponding cluster, followed by re-binning as described earlier. The resulting 5 MAGs vary in size from 4 to 5.8 Mb, which is within the expected range for the E. coli genome [104] (Fig. S3D). Each community model contained only one E. coli strain model (the one corresponding to the cluster that the sample belonged to), preventing overrepresentation of E. coli in community-level analyses. Although tools such as dRep can dereplicate genomes on the basis of ANI and quality metrics [105], our custom approach allowed us to optimize MAG quality specifically for metabolic modeling purposes by controlling how reads were pooled across samples. This approach enabled us to address strain heterogeneity in E. coli while maintaining consistent representation of other species across samples. Future studies with more extensive strain variation might benefit from dRep’s automated clustering algorithms, particularly when working with larger datasets from diverse environments. Additionally, while we assessed our MAGs based on completeness and contamination thresholds most relevant for metabolic modeling, we note that the majority of the produced MAGs does not contain the full complement of 5S, 16S, and 23S rRNA genes required by the strictest MIMAG criteria for “high-quality” classification [106], a common limitation in short-read metagenomic studies that does not significantly impact the recovery of metabolic genes necessary for our analyses. While most species benefited from sample-specific assembly followed by read pooling, five species yielded higher quality MAGs from the initial co-assembly of all 33 samples, based on completeness, contamination, and N50 metrics. These species were Bacteroides fragilis, Anaerostipes butyraticus, Alistipes excrementavium, Limihabitans stercoravium, and Caccovivens sp944345705. The characteristics and classification of the 237 reconstructed MAGs are provided in Table S5. Model reconstruction and curation The MAGs with a minimum relative abundance of 0.5% in each sample underwent reconstruction using gapseq v1.2 (commit 3e2d9393, Sequence DB md5sum: 17e92c9) with default settings [107], resulting in a total of 233 metabolic models. This 0.5% threshold was selected to balance comprehensive community representation with the computational constraints of metabolic modeling. Each sample was represented by 5 to 33 unique MAGs meeting our quality criteria. While we expect the samples to include more taxa, our analysis focused only on those meeting both our abundance threshold (≥ 0.5% relative abundance) and quality criteria (≥ 50% completeness, ≤ 10% contamination) to ensure reliable metabolic predictions. This approach utilizes only assembled genomes rather than the full metagenomic datasets, as genome-scale metabolic modeling requires defined metabolic networks for each organism. Since the specific dietary composition used as chicken feed was not disclosed in the original study, the nutritional input for the models was formulated based on a typical chicken corn-based diet [108] by translating molecular constituents into corresponding metabolites using vmh.org, foodb.ca, and feedtables.com resources [109–111]. Due to limited information on fiber in the chicken diet, cellulose was utilized as a primary source of fiber according to available literature [112, 113]. The absorption and digestion of metabolites in the upper gut compartments were taken into account when calculating the dietary input using published rates of nutrient disappearance along the distal ileum in poultry [114–119]. Specifically, we applied literature–derived fractional absorption factors to each nutrient class (e.g., starch concentrations were reduced by ~ 97% to reflect small-intestinal digestion before material enters the ceca). Where compound-specific data were unavailable, we adopted conservative estimates based on closely related nutrients. The resulting dietary input table containing the complete list of all metabolites with their concentrations is available in the GitHub repository (see the “Data availability” section). Additionally, urea and mucus were introduced to depict the cecal intestinal setting. Subsequently, this resulting nutritional composition served as an input medium for the gap-filling step via gapseq. To ensure microbial degradation capabilities towards dietary components, profiling of carbohydrate-active enzymes (CAZymes) encoded within metagenomes used for model reconstruction was performed using dbCAN3 [120], which is based on the CAZy database—an expert-curated and routinely updated resource. Predicted CAZymes were filtered using an e-value threshold (> 1e − 50) and required a minimum 90% coverage with enzyme(s) assigned EC number(s). Predicted EC numbers along with substrates were mapped to respective reactions within the modelSEED database and incorporated into models through gapseq adapt. Additionally, metabolic models of the species known as SCFA producers were examined for the presence of fermentation pathways documented in the literature; missing pathways were added to the models. Quantitative PCR Total DNA was extracted from cecal content using QIAamp Fast DNA Stool mini kit with the following manufacturer’s protocol with an additional bead-beating step for 6.0 m/s for 60 s (FastPrep-24 5G instrument, MP Biomedicals, Inc). qPCR of cecal digesta was performed using PercfeCTa SYBR green Supermix (Quantabio, Beverly, MA) on an ABI StepOne real-time system (Applied Biosystems, Foster City, CA) for the qPCR assays. qPCR condition was set at 95 °C for 3 min, 40 cycles of 95 °C for 10 s, and 60 °C for 30 s. Gene standards were generated through PCR amplification of the 16S rRNA fragment with the same pair of primers used for qPCR (forward CGGYCCAGACTCCTACGGG, reverse TTACCGCGGCTGCTGGCAC [121]). Concentrations of the amplified gene fragment were determined using a Quant-iT PicoGreen dsDNA assay kit (Invitrogen, Waltham, MA) and were used to create a standard curve. Copy numbers of the gene fragment were determined using the relative standard curve method and normalized to the weight of cecal content used for bacterial DNA extraction. Because each cecal community comprises many taxa with unknown or variable 16S operon counts, we did not apply any correction for copy-number variation; values therefore represent total 16S gene abundance at the community level. Set up of in silico ceca and simulation of sample-specific community models Microbial communities were simulated using the BacArena software[32], which combines Flux Balance Analysis and individual-based modeling. BacArena represents each organism individually on a two-dimensional grid to model a spatial environment. Metabolites can be exchanged between individuals and diffuse in the environment, allowing for the dynamic update of the state of each microbe and its surroundings. The initial microbial community was seeded on a 100 × 100 cell grid. To mimic the flow of luminal contents in and out of the cecum while maintaining a relatively stable bacterial population, the following approach was implemented. Multiple studies have shown that the cecal contents are released only 2–3 times per day [122, 123], which is a key feature for realistic simulation, considering the cecum’s role as a major hub for bacterial fermentation and SCFA production. At each hourly iteration, metabolites representing the metabolic contents from the upstream compartment were added to the grid and evenly distributed between the grid cells, simulating environmental mixing and allowing access to dietary and bacterial-derived metabolites to all cells within the simulated environment. Every fourth iteration, a fraction of the cecum, including both metabolites and microbes, was randomly removed to mimic emptying of the cecum. The number of grid cells selected for removal of contents was determined to be 25% of the total to maintain realistic fluctuations in bacterial density and a generally consistent community size. To reflect the physiological conditions of the chicken ceca, oxygen concentrations in simulations were set to 1e–08 mM to represent the nearly anaerobic environment of the chicken ceca, along with a pH of 7 [124, 125]. The absolute abundances of microbial communities were measured via quantitative PCR (qPCR) assay, yielding copy numbers per gram of digesta ranging from 1.62e+09 to 3.02e+11. To reduce skewness and ensure smaller values were distinguishable while preventing larger values from dominating, abundance values were log-transformed. To preserve these differences in bacterial density between samples while making them computationally tractable, these values were then rescaled to range from 1000 to 5000 total individuals per community in each sample. This scaling approach maintains the proportional differences in bacterial density between samples, which is important for comparing metabolic outputs such as SCFA production across samples. At the same time, it ensures that even species with low abundances were represented by at least five individuals on the grid while also allowing space for the community to grow and move within a grid of 10,000 cells. The communities were simulated in five replicates during 16 iterations, each representing an hour, to mimic the day of feeding. Comparison of functional capacities of community models Sets of unique reactions and pathways within each sample-specific microbiome were extracted from the corresponding metabolic models comprising the sample. Binary matrices indicating the presence or absence of reactions and pathways in each sample were created for subsequent Jaccard distance calculations. Additionally, to quantify the metabolic potential within each sample while accounting for the varying contributions of different species, we summarized the reaction and pathway counts within each sample. Each reaction or pathway count was weighted by the relative abundance of the species contributing to that reaction or pathway. The resulting incidence matrices provided detailed information on the pathway and reaction abundance in the samples. Dissimilarity matrices were constructed from the incidence matrices described using vegdist function from the vegan R-package, and Principal Coordinate Analysis was performed using cmdscale function from the stats R-package. Functional similarity networks detailing overlap between sample-specific microbiomes were constructed using a mutual information score calculated via mutinformation function from the infotheo R-package based on the metabolic reaction content of the samples. Networks were visualized using Cytoscape v.3.10.3 [126]. Community metabolic dissimilarity (CMD) score was calculated as the average of pairwise metabolic dissimilarities based on EC numbers within the community, adjusted for the community size. Specifically, the Jaccard distance matrix for a sample was calculated from a presence/absence matrix where the columns represent species and the rows represent EC numbers, using the vegdist function from the vegan R-package. The CMD score was then computed as a mean of the dissimilarity matrix, with further log-adjustment to account for the diminishing contribution of additional species in larger communities. Evaluation of cross-feeding and competition in in silico microbiomes To evaluate the extent of metabolic competition within Bacteroides-rich and Bacteroides-depleted microbial communities, the metabolic resource overlap (MRO) between species was quantified as previously described [43]. MRO is a metric that estimates the extent of overlap between the growth requirements of two or more metabolic models [43]. As such, it quantifies the potential for species to compete for a predetermined set of compounds in the media. Cross-feeding interactions within microbial communities were investigated using metabolic reaction fluxes. We filtered negligible fluxes to retain absolute total flux values above 1e–06 mmol per hour per gram of dry weight (mmol/gDW/hour) and categorized species as metabolite producers or consumers. By merging producer and consumer datasets on shared metabolites, cross-feeding interactions were summarized by counting and normalizing interaction occurrences across replicates within each category (based on Bacteroides presence). Key cross-feeders were identified using a threshold of at least 10 interactions involving the exchange of 12 or more distinct metabolites. These thresholds were selected to focus on taxa with substantial and consistent metabolic exchanges while filtering out weaker or sporadic interactions for clarity of visualization and interpretation. Cross-feeding was quantified by retrieving all metabolite exchanges (import and export fluxes) during simulations and calculating the total number of interactions (cross-feeding pairs) in each sample. A cross-feeding coefficient was introduced as a metric of cross-feeding activity within a community, defined as the number of total cross-feeding pairs formed divided by all possible combinations of species pairs in the community. Interaction data were visualized using the ComplexHeatmap R-package. Next, the total production and consumption for each metabolite were calculated over the course of the simulation. The contribution of each species to the production and consumption was determined as a percentage of the total production/consumption. Exchanged metabolites were filtered to include only nondietary metabolites, which are exclusively produced by bacteria. The pairwise interactions involving non-dietary metabolites summarized across samples are visualized as a heatmap in Fig. S6A, with raw data used to plot the heatmap provided in Table S3. To focus on metabolic niche adjustments involving keystone taxa, metabolites were further filtered based on the following criteria after completing all primary statistical comparisons between HB and NB communities, with this filtering applied specifically for visualization and detailed cross-feeding analysis in Fig. 5: (1) they are predominantly produced or consumed by keystone taxa, defined as those where at least 30% of produced metabolite is attributed to the production or consumption by a single keystone taxon; (2) to ensure the metabolite is involved in cross-feeding, at least 1% of the produced amount must be consumed by at least one other taxon; and (3) metabolites produced in highest quantities by the same taxon in both HB and NB communities were excluded to emphasize shifts in metabolic roles. These metabolite filtering criteria helped identify metabolites with significant keystone taxa influence that were genuinely exchanged between community members rather than theoretically available in trace amounts. This filtering process resulted in the identification of 22 metabolites featuring different cross-feeding relationships in HB and NB communities. Of these, 18 were selected for downstream visualization. Simple sugars d-gluconate, d-ribose, and d-arabinose were excluded due to their relatively low rates of production; l-lactate was excluded because its production and consumption patterns significantly overlapped with those of d-lactate, so only d-lactate is shown. For each of these metabolites, the relative contributions of the top three producers and consumers were visualized (Fig. 5). Sankey diagrams were generated using the networkD3 R-package. To highlight major contributors, a minimum contribution threshold of 5% was used to filter taxa for visualization. To investigate differences in metabolite production and consumption between HB and NB microbial communities, the fluxes of exchanged metabolites were summarized within each community at every hour of the simulation and normalized by the total community size at that time point. Subsequently, the community fluxes (normalized by abundance) were aggregated over the 16-h simulation period, and differences between the categories were evaluated using the Wilcoxon rank-sum statistical test. For graphical representation, metabolic fluxes were log2-transformed. Statistical analyses All statistical analyses were performed in R v4.2.3. Significant differences for numbers of reactions and pathways, metabolite concentrations, metabolic fluxes, community metabolic diversity, and cross-feeding activity coefficients were determined by non-parametric Mann–Whitney U-test (Wilcoxon rank-sum test). FDR correction (Benjamini–Hochberg procedure) was applied for all statistical tests with multiple comparisons. Spearman’s rank correlation coefficients and associated p values were calculated between measured and predicted SCFA concentrations. Given E. coli’s high degree of connectivity, we explored its influence on metabolic exchanges within microbial communities. To evaluate how the presence or absence of this pivotal taxon determines the degree of cross-feeding activity, we constructed a linear model incorporating community size and the abundance of E. coli alongside key metabolic hubs: A. butyraticus, L. crispatus, and B. fragilis using the lm function in R. The response variable was the cross-feeding coefficient (CrossFeeding_coef), which quantifies the degree of metabolic interactions within the community. Predictor variables included categorical levels of E. coli abundance (Ecoli_category: “NoEcoli,” “LowEcoli” (< 5% relative abundance), “HighEcoli” (> 5% relative abundance)), Bacteroides presence or absence that defines the community category (Bacteroides_presence: “B” or “HB”), the presence of key taxa such as Anaerostipes butyraticus (Abutyraticus_presence) and Lactobacillus crispatus (Lcrispatus_presence), and community size (comm_size). Interaction terms between E. coli abundance and the other predictors were included to capture potential synergistic or antagonistic effects on cross-feeding. The model was specified as follows: CrossFeeding_coef ~ Ecoli_category:(Bacteroides_presence + Abutyraticus_presence + Lcrispatus_presence) + comm_size. The model summary is provided in TableS3. All statistical analyses employed a significance threshold of p-value < 0.05. Supplementary Information Supplementary Material 1. Supplementary Figure 1. Relative abundances of the common taxa across samples. This heatmap illustrates the relative abundances of the top 52 most common taxa across 33 cecal samples, categorized into high Bacteroides (HB, in gold) and no Bacteroides (NB, in green) groups. Taxa present in at least four samples are shown, with the relative abundance represented by the color intensity of each tile.Supplementary Material 2. Supplementary Figure 2. Distribution of enzyme commission numbers (ECs) across metabolic models. The columns of the heatmap represent individual ECs, while the rows correspond to metabolic models representing different taxa. ECs are annotated using KEGG pathway designations, indicated in two layers to reflect primary and secondary pathway associations. For ECs lacking KEGG pathway annotation, MetaCyc pathway annotation was utilized when available. For ECs mapped to 3 or more major pathway categories, “Multiple superpathways” category was assigned. Each row represents a different taxa, grouped by their presence in HB only, NB only, or both; color-coded.Supplementary Material 3. Supplementary Figure 3. Characteristics of 237 reconstructed MAGs. A-C. Boxplots showing beta-dispersion analysis results - the distances from each sample to the centroid of its respective group (HB or NB). Lower values indicate samples within that group are more similar to each other, while higher values indicate greater within-group variability. The significant differences in dispersion between HB and NB communities (particularly for reaction and pathway abundance) reveal that NB communities exhibit more variable metabolic profiles than HB communities, suggesting that Bacteroides presence may constrain metabolic variability. Distances were calculated from: (A) PCoA based on reaction content (top: Jaccard distances for presence/absence; bottom: Bray-Curtis dissimilarities for abundance-weighted data); (B) PCoA based on pathway content (top: Jaccard distances; bottom: Bray-Curtis dissimilarities); and (C) taxonomic composition at the family level. (). Significance determined using beta-dispersion (TukeyHSD) test. * - p-value < 0.05, ** - p-value <0.01, *** - p-value < 0.001. D. Major characteristics of 237 MAGs reconstructed in this study: completeness, contamination, N50 and genome size.Supplementary Material 4. Supplementary Figure 4. SCFA reaction fluxes at the species-level. Cumulative fluxes over 16 hours of simulation for acetate, propionate and butyrate, with samples (x-axis) ordered by the corresponding experimentally measured concentrations – from smallest to largest (from left to right). Colors correspond to the specie contributing to the production or consumption. Positive fluxes correspond to production, while negative fluxes indicated consumption.Supplementary Material 5. Supplementary Figure 5. Influence of species-specific metabolic potential and key taxa presence on cross-feeding activity. A-B. Correlation plots showing the relationship between the average number of unique EC numbers presented in a specie within a sample and (A) the total number of interactions formed by a specie normalized by the number of samples where a specie is present (Pearson) (B) the number of samples where the specie is present (Pearson). Red line is fitted linear regression, and gray area indicates the 95% confidence interval. Each data point represents a species. Only species with at least 10 cross-feeding pairs involving at least 12 metabolites are shown. Selected species (present in at least two samples, above the threshold of 230 total interactions per sample and average number of unique EC per sample > 5) have labels that are color-coded corresponding to the taxonomic family they belong to. C. Effect of E. coli presence across both community categories (HB - High Bacteroides; NB - no Bacteroides) on cross-feeding activity, measured by the cross-feeding coefficient (CFC). Each dot corresponds to a CFC calculated for a single sample. Points and lines are color-coded based on E. coli abundance: red represents communities with high E. coli abundance (>5%), blue represents communities with low E. coli abundance (<5%), and gray represents communities where E. coli is absent. In the NB group, higher E. coli abundance is associated with a notable increase in cross-feeding activity compared to low or absent E. coli, suggesting a significant role for E. coli in enhancing cross-feeding, particularly in the absence of Bacteroides. Conversely, in HB communities, E. coli presence appears to have a more muted effect on cross-feeding, especially in low abundance. D-E. Interaction effects of E. coli and A. butyraticus (D)or L. crispatus (E)on cross-feeding activity, measured by the CFC, across categories. Different colors represent distinct interaction scenarios: E. coli & A. butyraticus (D)/L. crispatus (E) absent (gray), E. coli absent & A. butyraticus (D)/L. crispatus (E) present (cyan), low-abundance E. coli & A. butyraticus (D)/L. crispatus (E) absent (blue), low-abundance E. coli & A. butyraticus (D)/L. crispatus (E) present (orange), high-abundance E. coli & A. butyraticus (D)/L. crispatus (E) absent (red), and high-abundance E. coli & A. butyraticus (D)/L. crispatus (E) present (purple). Each point represents an average CFC for samples within each interaction scenario. NB communities with high E. coli abundance and without A. butyraticus (solid red line, D) show the highest cross-feeding activity. When A. butyraticus is present (dotted lines, D), the interaction patterns change: the cross-feeding coefficient decreases slightly in communities with high E. coli, but it increases in communities with low or no E. coli. This suggests that A. butyraticus may partly compensate for the absence or lower abundance of E. coli, contributing to cross-feeding, particularly in NB communities. The interaction between A. butyraticus and E. coli seems to reveal a potentially competitive relationship where the presence of both may reduce cross-feeding, while the presence of either one alone promotes cross-feeding in different ways. Presence of L. crispatus is associated with increased cross-feeding activity in NB compared to HB communities (E), regardless of E. coli abundance, while absence of L. crispatus doesn’t increase significantly the cross-feeding activity in NB when E. coli levels are low. This points to an impact of antagonistic relationship between L. crispatus and Bacteroides rather than potential competitive relationship between L. crispatus and E. coli. Supplementary Material 6. Supplementary Figure 6. Non-dietary metabolite exchanges across all 227 taxa in HB and NB communities. A. The heatmap displays the normalized total interactions involving non-dietary metabolites exchanged between 227 taxa across 33 chicken cecum samples. Interactions were normalized by the number of samples in which each cross-feeding pair was present and are color-coded by interaction frequency. Taxa are color-coded based on their taxonomic families, interaction frequencies are indicated by color intensity. Taxa on y-axis are grouped by their presence in HB only, NB only, or both; color-coded. Distinct cliques of interacting species are highlighted in red boxes (A-H), which were selected based on visual inspection of the hierarchically clustered heatmap to identify the largest and most distinct clusters of taxa with similar interaction patterns. B. The upset plot summarizes the unique and shared non-dietary metabolites exchanged within the investigated cliques (A-H) highlighted in panel A. Each bar represents the number of metabolites exchanged, with intersections indicating metabolites shared between multiple cliques. Some metabolite names have been shortened or modified for clarity in the figure. The following abbreviations are used: 2-Oxoglut for 2-Oxoglutarate, Ddca for dodecanoic acid, L-Asp for L-Aspartate, Glu for Glutamate, D-Lact and L-Lact for D- and L-Lactate, respectively, NAD for nicotinamide adenosine dinucleotide, Propion for propionate, Pyro-P for pyrophosphate, and Putrsc for putrescine.Supplementary Material 7. Supplementary Table 1. EC content in reconstructed metagenome assembled genomes annotated with KEGG and MetaCyc (related to Supplementary Figure 2).Supplementary Material 8. Supplementary Table 2. Normalized cross-feeding interactions between top cross-feeding taxa in HB and NB communities (related to Figure 4 A).Supplementary Material 9. Supplementary Table 3. Linear model summary: impact of key metabolic hubs presence on cross-feeding activity. The response variable is the cross-feeding coefficient, which quantifies the degree of metabolic interactions within the community. Predictor variables included categorical levels of E. coli abundance (Ecoli_category:"NoEcoli","LowEcoli” (<5% relative abundance),"HighEcoli (>5% relative abundance)"), Bacteroides presence or absence that defines the community category (Bacteroides:"NB"or"HB"), the presence of key taxa such as Anaerostipes butyraticus (Abutyraticus_presence) and Lactobacillus crispatus (Lcrispatus_presence), and community size (comm_size). Supplementary Material 10. Supplementary Table 4. Non-dietary metabolite exchanges between taxa across HB and NB communities (related to Supplementary Figure 6). Supplementary Material 11. Supplementary Table 5. Detailed characteristics and classification of the 237 reconstructed MAGs. This table provides details on all metagenome-assembled genomes (MAGs) reconstructed from 33 chicken cecal samples. The table includes quality metrics (completeness, contamination, N50) for each MAG along with taxonomic classification, abundance data, and sample category (HB: High Bacteroides; NB: No Bacteroides). The"binID used for metabolic reconstruction"column indicates which specific MAG was used to create the metabolic model for each species. Our naming convention reflects the assembly strategy used: i. SpecieName_pooled: For most species present across multiple samples, reads were pooled from all samples where the species was detected, reassembled, and binned to generate a single high-quality MAG. ii. Ecoli_cluster1 through Ecoli_cluster5: Due to E. coli's high prevalence (23 samples) and strain heterogeneity, we clustered the original E. coli bins into 5 groups based on average nucleotide identity (ANI). For each cluster, reads were pooled from samples within that cluster to generate 5 distinct strain-specific MAGs. Each E. coli strain model was only included in community models for samples belonging to that specific cluster. iii. Sample_SpecieName: For species unique to a single sample, MAGs were derived directly from individual sample assemblies. iv. SpecieName_coassembly: Five species (Bacteroides fragilis, Anaerostipes butyraticus, Alistipes excrementavium, Limihabitans stercoravium, and Caccovivens sp944345705) yielded higher quality MAGs from co-assembly of all 33 samples rather than from pooled reads.Supplementary Material 12. Supplementary Table 6. Sequencing statistics of the 33 raw metagenomic DNA datasets used in this study. The table includes sample identifiers, total number of reads, total number of bases (i.e., dataset size), minimum, average, and maximum read lengths. Supplementary Material 1. Supplementary Figure 1. Relative abundances of the common taxa across samples. This heatmap illustrates the relative abundances of the top 52 most common taxa across 33 cecal samples, categorized into high Bacteroides (HB, in gold) and no Bacteroides (NB, in green) groups. Taxa present in at least four samples are shown, with the relative abundance represented by the color intensity of each tile. Supplementary Material 2. Supplementary Figure 2. Distribution of enzyme commission numbers (ECs) across metabolic models. The columns of the heatmap represent individual ECs, while the rows correspond to metabolic models representing different taxa. ECs are annotated using KEGG pathway designations, indicated in two layers to reflect primary and secondary pathway associations. For ECs lacking KEGG pathway annotation, MetaCyc pathway annotation was utilized when available. For ECs mapped to 3 or more major pathway categories, “Multiple superpathways” category was assigned. Each row represents a different taxa, grouped by their presence in HB only, NB only, or both; color-coded. Supplementary Material 3. Supplementary Figure 3. Characteristics of 237 reconstructed MAGs. A-C. Boxplots showing beta-dispersion analysis results - the distances from each sample to the centroid of its respective group (HB or NB). Lower values indicate samples within that group are more similar to each other, while higher values indicate greater within-group variability. The significant differences in dispersion between HB and NB communities (particularly for reaction and pathway abundance) reveal that NB communities exhibit more variable metabolic profiles than HB communities, suggesting that Bacteroides presence may constrain metabolic variability. Distances were calculated from: (A) PCoA based on reaction content (top: Jaccard distances for presence/absence; bottom: Bray-Curtis dissimilarities for abundance-weighted data); (B) PCoA based on pathway content (top: Jaccard distances; bottom: Bray-Curtis dissimilarities); and (C) taxonomic composition at the family level. (). Significance determined using beta-dispersion (TukeyHSD) test. * - p-value < 0.05, ** - p-value <0.01, *** - p-value < 0.001. D. Major characteristics of 237 MAGs reconstructed in this study: completeness, contamination, N50 and genome size. Supplementary Material 4. Supplementary Figure 4. SCFA reaction fluxes at the species-level. Cumulative fluxes over 16 hours of simulation for acetate, propionate and butyrate, with samples (x-axis) ordered by the corresponding experimentally measured concentrations – from smallest to largest (from left to right). Colors correspond to the specie contributing to the production or consumption. Positive fluxes correspond to production, while negative fluxes indicated consumption. Supplementary Material 5. Supplementary Figure 5. Influence of species-specific metabolic potential and key taxa presence on cross-feeding activity. A-B. Correlation plots showing the relationship between the average number of unique EC numbers presented in a specie within a sample and (A) the total number of interactions formed by a specie normalized by the number of samples where a specie is present (Pearson) (B) the number of samples where the specie is present (Pearson). Red line is fitted linear regression, and gray area indicates the 95% confidence interval. Each data point represents a species. Only species with at least 10 cross-feeding pairs involving at least 12 metabolites are shown. Selected species (present in at least two samples, above the threshold of 230 total interactions per sample and average number of unique EC per sample > 5) have labels that are color-coded corresponding to the taxonomic family they belong to. C. Effect of E. coli presence across both community categories (HB - High Bacteroides; NB - no Bacteroides) on cross-feeding activity, measured by the cross-feeding coefficient (CFC). Each dot corresponds to a CFC calculated for a single sample. Points and lines are color-coded based on E. coli abundance: red represents communities with high E. coli abundance (>5%), blue represents communities with low E. coli abundance (<5%), and gray represents communities where E. coli is absent. In the NB group, higher E. coli abundance is associated with a notable increase in cross-feeding activity compared to low or absent E. coli, suggesting a significant role for E. coli in enhancing cross-feeding, particularly in the absence of Bacteroides. Conversely, in HB communities, E. coli presence appears to have a more muted effect on cross-feeding, especially in low abundance. D-E. Interaction effects of E. coli and A. butyraticus (D)or L. crispatus (E)on cross-feeding activity, measured by the CFC, across categories. Different colors represent distinct interaction scenarios: E. coli & A. butyraticus (D)/L. crispatus (E) absent (gray), E. coli absent & A. butyraticus (D)/L. crispatus (E) present (cyan), low-abundance E. coli & A. butyraticus (D)/L. crispatus (E) absent (blue), low-abundance E. coli & A. butyraticus (D)/L. crispatus (E) present (orange), high-abundance E. coli & A. butyraticus (D)/L. crispatus (E) absent (red), and high-abundance E. coli & A. butyraticus (D)/L. crispatus (E) present (purple). Each point represents an average CFC for samples within each interaction scenario. NB communities with high E. coli abundance and without A. butyraticus (solid red line, D) show the highest cross-feeding activity. When A. butyraticus is present (dotted lines, D), the interaction patterns change: the cross-feeding coefficient decreases slightly in communities with high E. coli, but it increases in communities with low or no E. coli. This suggests that A. butyraticus may partly compensate for the absence or lower abundance of E. coli, contributing to cross-feeding, particularly in NB communities. The interaction between A. butyraticus and E. coli seems to reveal a potentially competitive relationship where the presence of both may reduce cross-feeding, while the presence of either one alone promotes cross-feeding in different ways. Presence of L. crispatus is associated with increased cross-feeding activity in NB compared to HB communities (E), regardless of E. coli abundance, while absence of L. crispatus doesn’t increase significantly the cross-feeding activity in NB when E. coli levels are low. This points to an impact of antagonistic relationship between L. crispatus and Bacteroides rather than potential competitive relationship between L. crispatus and E. coli. Supplementary Material 6. Supplementary Figure 6. Non-dietary metabolite exchanges across all 227 taxa in HB and NB communities. A. The heatmap displays the normalized total interactions involving non-dietary metabolites exchanged between 227 taxa across 33 chicken cecum samples. Interactions were normalized by the number of samples in which each cross-feeding pair was present and are color-coded by interaction frequency. Taxa are color-coded based on their taxonomic families, interaction frequencies are indicated by color intensity. Taxa on y-axis are grouped by their presence in HB only, NB only, or both; color-coded. Distinct cliques of interacting species are highlighted in red boxes (A-H), which were selected based on visual inspection of the hierarchically clustered heatmap to identify the largest and most distinct clusters of taxa with similar interaction patterns. B. The upset plot summarizes the unique and shared non-dietary metabolites exchanged within the investigated cliques (A-H) highlighted in panel A. Each bar represents the number of metabolites exchanged, with intersections indicating metabolites shared between multiple cliques. Some metabolite names have been shortened or modified for clarity in the figure. The following abbreviations are used: 2-Oxoglut for 2-Oxoglutarate, Ddca for dodecanoic acid, L-Asp for L-Aspartate, Glu for Glutamate, D-Lact and L-Lact for D- and L-Lactate, respectively, NAD for nicotinamide adenosine dinucleotide, Propion for propionate, Pyro-P for pyrophosphate, and Putrsc for putrescine. Supplementary Material 7. Supplementary Table 1. EC content in reconstructed metagenome assembled genomes annotated with KEGG and MetaCyc (related to Supplementary Figure 2). Supplementary Material 8. Supplementary Table 2. Normalized cross-feeding interactions between top cross-feeding taxa in HB and NB communities (related to Figure 4 A). Supplementary Material 9. Supplementary Table 3. Linear model summary: impact of key metabolic hubs presence on cross-feeding activity. The response variable is the cross-feeding coefficient, which quantifies the degree of metabolic interactions within the community. Predictor variables included categorical levels of E. coli abundance (Ecoli_category:"NoEcoli","LowEcoli” (<5% relative abundance),"HighEcoli (>5% relative abundance)"), Bacteroides presence or absence that defines the community category (Bacteroides:"NB"or"HB"), the presence of key taxa such as Anaerostipes butyraticus (Abutyraticus_presence) and Lactobacillus crispatus (Lcrispatus_presence), and community size (comm_size). Supplementary Material 10. Supplementary Table 4. Non-dietary metabolite exchanges between taxa across HB and NB communities (related to Supplementary Figure 6). Supplementary Material 11. Supplementary Table 5. Detailed characteristics and classification of the 237 reconstructed MAGs. This table provides details on all metagenome-assembled genomes (MAGs) reconstructed from 33 chicken cecal samples. The table includes quality metrics (completeness, contamination, N50) for each MAG along with taxonomic classification, abundance data, and sample category (HB: High Bacteroides; NB: No Bacteroides). The"binID used for metabolic reconstruction"column indicates which specific MAG was used to create the metabolic model for each species. Our naming convention reflects the assembly strategy used: i. SpecieName_pooled: For most species present across multiple samples, reads were pooled from all samples where the species was detected, reassembled, and binned to generate a single high-quality MAG. ii. Ecoli_cluster1 through Ecoli_cluster5: Due to E. coli's high prevalence (23 samples) and strain heterogeneity, we clustered the original E. coli bins into 5 groups based on average nucleotide identity (ANI). For each cluster, reads were pooled from samples within that cluster to generate 5 distinct strain-specific MAGs. Each E. coli strain model was only included in community models for samples belonging to that specific cluster. iii. Sample_SpecieName: For species unique to a single sample, MAGs were derived directly from individual sample assemblies. iv. SpecieName_coassembly: Five species (Bacteroides fragilis, Anaerostipes butyraticus, Alistipes excrementavium, Limihabitans stercoravium, and Caccovivens sp944345705) yielded higher quality MAGs from co-assembly of all 33 samples rather than from pooled reads. Supplementary Material 12. Supplementary Table 6. Sequencing statistics of the 33 raw metagenomic DNA datasets used in this study. The table includes sample identifiers, total number of reads, total number of bases (i.e., dataset size), minimum, average, and maximum read lengths.