An atlas of colonization factors in the human gut microbiome reveals ecological strategies and inflammatory bowel disease signatures

Abstract

Long-term residence in the gut enables microbes to interact with the host and influence intestinal health. However, many microbiome studies focus on taxonomic profiles or broad metabolic pathway annotations and provide limited insight into the conserved genes that support microbial residence. Here we show that colonization factors (CFs), defined as gut-enriched genes associated with microbial residence, offer a colonization-centered functional framework for profiling the human gut microbiome. By mapping 79 CF families across 289,231 surveyed microbial genomes, we identify more than seven million CF homologs and reveal their widespread distribution, phylogenetically structured organization and functional stratification into three putative lineage-associated colonization strategies centered on metabolism, stress resistance and microbial communication. Applying this framework to 3,666 metagenomic and metatranscriptomic samples from 10 inflammatory bowel disease (IBD) cohorts, we find that disease-associated dysbiosis is accompanied by recurrent remodeling of CF repertoires. These alterations can be traced to specific colonization-associated functions and microbial carrier species, and are captured by a compact feature panel that discriminates disease status within cohorts. These results establish CF profiling as a mechanism-oriented approach for interpreting the ecological organization of the gut microbiome and prioritizing colonization-associated targets for future disease monitoring and intervention studies.

Here, Meng et al. establish a human gut microbiome atlas of colonization factors, revealing distinct ecological strategies used by microbes to persist in the intestine, and show that inflammatory bowel disease is associated with remodeling of colonization-associated functions and microbial carriers.

Introduction

The human gut microbiota is a complex ecological community that has co-evolved with its host over millions of years. In this symbiotic system, commensals provide essential metabolic functions, contribute to the integrity of the intestinal epithelial barrier, regulate immune homeostasis, and protect against pathogen invasion^1–3, whereas potentially pathogenic members may produce harmful metabolites or promote inflammation^3–5. These effects depend not only on which microbes are present but also on how they establish residence, compete, and interact within the intestinal environment.

Microbial colonization is a biological prerequisite for these host–microbe interactions. To influence gut ecology or host physiology, microorganisms must first survive gastrointestinal transit, adapt to nutrient and stress gradients, compete with resident microbes, and persist within intestinal niches^6–8. Experimental approaches, including transposon-insertion sequencing (Tn-Seq) and hybrid selection RNA sequencing (hsRNA-Seq), have identified colonization factors (CFs) in selected pathogens and model commensals^9–17. These studies have demonstrated that colonization depends on a range of genetic functions related to nutrient acquisition, stress tolerance, adhesion, signaling, and niche adaptation. However, experimental CF discovery remains limited to a small number of tractable species and does not yet provide a microbiome-wide view of colonization-associated functional potential.

A recent genotype–habitat association study by Liu et al. provided an important foundation for extending CF discovery beyond experimentally tractable species¹⁸. By comparing approximately 9,475 high-quality microbial genomes derived from human and murine intestines as well as environmental sources, they identified 27,096 colonization-associated genes and clustered them into 79 CF families. This work established a computational framework for the large-scale discovery of candidate genetic determinants of mammalian gut colonization. However, several key questions remain unresolved. First, the distribution of these CF families across human gut microbial genomes has not been systematically mapped, nor have their abundance and expression patterns been comprehensively characterized in human gut metagenomes and metatranscriptomes. Second, it remains unclear whether combinations of CF families form biologically interpretable colonization-associated strategies across microbial lineages. Third, the relevance of CF repertoires to disease-associated community remodeling remains largely unexplored.

Addressing these questions may provide a colonization-centered dimension that is largely absent from current human gut microbiome research, which predominantly relies on taxonomic profiling or broad functional annotation¹⁹. These approaches have been highly successful in describing dysbiosis, identifying disease-associated taxa, and cataloging metabolic pathway alterations, but they provide limited resolution of colonization biology. Taxonomic profiles indicate which microbial lineages are altered, whereas conventional functional profiles capture broad metabolic potential that is often difficult to connect directly to microbial residence, ecological fitness, and host interaction. Thus, colonization-associated functions represent an underexplored layer linking microbial identity, ecological strategy, and disease-associated community remodeling.

Inflammatory bowel disease (IBD) provides a relevant disease context for applying a CF-centered framework. IBD is characterized by chronic intestinal inflammation and consistently observed gut microbial dysbiosis, including reduced diversity, expansion of facultative anaerobes, depletion of commensal lineages, and altered microbial metabolism^20,21. These inflammatory and ecological disturbances are likely to reshape the selective pressures governing microbial persistence in the gut, yet the colonization-associated functions that support fitness within these altered niches remain insufficiently resolved. Because CFs represent a compact set of genes linked to gut residence, CF-family profiling may provide a mechanistically informed and biologically interpretable layer for connecting IBD-associated community shifts to colonization functions and their microbial carrier lineages.

Here, we constructed an integrated atlas of CF homologs across the human gut microbiome by mapping the 79 CF families defined by Liu et al.¹⁸ onto the Unified Human Gastrointestinal Genome (UHGG) and Unified Human Gastrointestinal Protein (UHGP) catalogs. We first mapped the CF homologs across 289,231 surveyed human gut microbial genomes and characterized their widespread distribution and lineage-associated organization across gut microbial species. We then used genome-encoded CF-family profiles to identify three putative, functionally distinct colonization-associated strategies and interpreted these strategies using genome-wide functional annotations. Finally, we applied this CF-centered framework to 3666 metagenomic and metatranscriptomic samples from 10 IBD cohorts. By integrating analyses of CF abundance, microbial carrier taxa, disease-status prediction, and transcriptional activity, we showed that CF repertoires provide a compact and biologically interpretable representation of IBD-associated dysbiosis, with changes traceable to specific colonization-associated functions and microbial carrier lineages. Together, this work establishes CF-family profiling as a colonization-centered framework for linking the ecology of microbial colonization to disease-associated community shifts and prioritizing colonization-associated targets for future mechanistic and translational studies.

Results

A human gut CF atlas reveals near-universal but lineage-differentiated colonization-associated repertoires

CFs represent genetic determinants that enable gut microbes to establish and persist within intestinal niches, thereby providing a colonization-centered functional perspective on gut microbial ecology. To characterize this functional layer at the microbiome scale, we constructed a human gut atlas of CF homologs by integrating a curated reference set of 27,096 CF proteins with the UHGP-90 and UHGG catalogs (Supplementary Data 1). Through sequence similarity searches, we identified 78,579 CF homologs in UHGP-90 that matched 7,536 reference CF proteins and were assigned to 79 curated CF families (Fig. 1a). Mapping these homologs to UHGG further identified a total of 7056,385 CF genes in 289,022 of the 289,231 surveyed genomes (99.9%), spanning 4,716 species (Fig. 1b). Among these, CF5 (multidrug and toxic compound extrusion [MATE] efflux family protein), CF0_0 (MATE efflux family protein), CF4 (cytidylate kinase-like family protein) and CF1 (anaerobic ribonucleoside-triphosphate reductase activating protein) contained the largest numbers of detected homologs. Species accumulation curves did not fully saturate, indicating that additional CF diversity may remain to be discovered with further expansion of gut microbial genome catalogs (Fig. 1c).

Fig. 1. A human gut CF atlas reveals near-universal but lineage-differentiated colonization potential.

a, b Doughnut plots showing the distribution of CF homologs identified in the UHGP-90 (a) and UHGG (b) catalogs. c Species accumulation curves showing the cumulative number of detected reference CF proteins with increasing numbers of sampled species. d Species-level phylogenetic tree of 4,716 microbial species carrying CF homologs. Branches are colored by phylum-level taxonomy. Grey highlights indicate clades corresponding to the 20 most species-rich genera. The colored middle ring denotes the 20 most species-rich families, and outer blue tracks indicate the presence of CF families across species. e, f Distributions of CF-family fractions across genus-level (e) and species-level (f) taxa, grouped by phylum-level taxonomy. The CF-family fraction represents the proportion of the 79 CF families detected in each taxon. Exact n values are shown in the group labels; n denotes the number of genus-level taxa in (e) and species-level taxa in (f). Each point represents one taxon. Boxes represent the interquartile range (IQR) extending from the first quartile to the third quartile, with the internal line indicating the median. Whiskers extend to the minimum and maximum values within 1.5 × IQR of the box boundaries. Points beyond the whiskers represent outliers. Statistical significance was assessed using the Kruskal–Wallis test.

To reduce gene-level annotation bias and characterize CF-family distributions across taxonomic groups, we generated a species-level CF-family presence/absence matrix and then aggregated it to higher taxonomic levels. At the phylum level, Firmicutes, Firmicutes_A, Actinobacteriota, Proteobacteria, Firmicutes_C and Fusobacteriota possessed the richest CF repertoires, each containing more than half of all CF families (Fig. 1d and Supplementary Fig. 1). Seven CF families were present in over 70% of the surveyed phyla, representing the most broadly conserved elements: CF7 (Cys-tRNA^Pro/Cys-tRNA^Cys deacylase), CF43 (rubrerythrin), CF17 (uncharacterized protein), CF5, CF11 (galactoside/maltose O-acetyltransferase), CF0_0 and CF38 (transcriptional regulator HxlR family protein).

At finer taxonomic scales, genus- and species-level analyses revealed pronounced heterogeneity in CF-family repertoire breadth (Kruskal–Wallis test, p < 0.001; Fig. 1e, f). Taxa within Firmicutes_A exhibited the greatest CF-family repertoire breadth, followed by Firmicutes, Actinobacteriota and Bacteroidota. Genera such as Clostridium, Enterocloster and Blautia_A, as well as species such as Clostridium butyricum, Enterocloster bolteae and Muricomes oroticus, harbored broader CF-family repertoires than taxa from many other lineages, suggesting broader colonization-associated functional potential. In contrast, taxa within Campylobacterota, Cyanobacteria and Verrucomicrobiota generally exhibited narrower CF-family repertoires. Together, these results establish a human gut CF atlas and show that CFs form a broadly distributed but lineage-distinct layer of microbial colonization potential.

CF repertoires retain detectable lineage-associated structure across microbial species

We next examined whether species-level CF repertoires were randomly distributed or associated with microbial lineages. Principal coordinates analysis (PCoA) of Jaccard dissimilarities based on CF-family presence/absence profiles showed taxonomic organization at both the phylum and family levels (Fig. 2a, b). Consistently, permutational multivariate analysis of variance (PERMANOVA) indicated that the proportion of variation in CF-family profiles explained by taxonomy increased from broad to finer taxonomic ranks, from 0.44% at the domain level to 27.67%, 33.44%, 47.47%, 59.83% and 74.91% at the phylum, class, order, family and genus levels, respectively (Fig. 2c). CF-family profile dissimilarity was also significantly associated with patristic distance, with a median Spearman’s r of 0.41 for bacterial species (p < 0.001) and an r of 0.76 for archaeal species (p < 0.001; Fig. 2d, e and Supplementary Fig. 2).

Fig. 2. CF repertoires retain detectable lineage-associated structure across microbial species.

a, b Principal coordinates analysis (PCoA) of species-level CF-family presence/absence profiles based on Jaccard dissimilarities. Points represent microbial species and are colored by phylum-level (a) or family-level (b) taxonomy. c Proportion of CF-family profile variation explained by different taxonomic ranks, measured as adjusted R² using permutational multivariate analysis of variance (PERMANOVA). P values were calculated using one-sided upper-tail permutation tests based on the pseudo-F statistic with 1000 permutations. d, e Associations between patristic distance and CF-family profile dissimilarity in bacterial (d) and archaeal (e) lineages. Spearman’s rank correlation was used to assess these associations. Two-sided P values were calculated by testing the null hypothesis that Spearman’s r was equal to zero. For bacteria, one representative 0.01% random subsample of nonredundant paired distances is shown; the main text reports the median Spearman’s r across 20 independent subsamples. For archaea, all nonredundant paired distances were used. f–k Matched randomization analyses comparing the proportion of variation in CF-like POG profiles explained by taxonomy with the corresponding proportions obtained from matched random POG backgrounds at the domain (f), phylum (g), class (h), order (i), family (j), and genus (k) levels. Red lines indicate observed CF-like POG statistics, and density curves represent empirical null distributions generated from matched random POG sets. Standardized effect size (SES) values indicate standardized deviations from the matched random background. Empirical one-sided P values were calculated as the proportion of matched random POG sets with statistics less than or equal to the observed statistic. l, m Matched randomization analyses of Mantel correlations between patristic distances and CF-like POG profile dissimilarities in bacterial (l) and archaeal (m) lineages. Red lines indicate observed CF-like POG statistics, and density curves represent empirical null distributions generated from matched random POG sets. Empirical one-sided P values were calculated as the proportion of matched random POG sets with statistics greater than or equal to the observed statistic.

We then evaluated whether this lineage association exceeded that expected for comparable protein-family backgrounds. To this end, we reconstructed a UHGP-90 protein-family background and generated matched random protein operational group (POG) sets with abundance, prevalence, and family-size distributions comparable to those of the CF-like POG set. Matched randomization showed that taxonomy explained less variation in CF-like POG profiles than in matched random POG profiles across taxonomic ranks, yielding negative standardized effect size (SES) values from the domain to genus levels (Fig. 2f–k). Similarly, although CF-like POG profiles retained detectable associations with phylogenetic distance, these associations were not significantly stronger than those observed for matched random POG sets in either bacteria (SES = −0.951, p = 0.817) or archaea (SES = 0.67, p = 0.101; Fig. 2l, m). Thus, CF repertoires retain detectable lineage-associated organization, but their taxonomic and phylogenetic structuring is not unusually strong relative to comparable protein-family backgrounds.

CF-family encoding patterns define three putative colonization-associated strategies across gut lineages

Because gut microbial lineages occupy distinct ecological niches and encode markedly different CF repertoires, we examined whether CF-family combinations could delineate alternative colonization-associated functional strategies. Based on the Jaccard dissimilarity matrix derived from species-level CF-family presence/absence profiles, partitioning around medoids (PAM) clustering identified three distinct microbial species clusters, designated Cluster-1, Cluster-2 and Cluster-3; k = 3 was supported by the Calinski–Harabasz criterion (Fig. 3a, b and Supplementary Data 2). To identify cluster-specific CF families, we calculated the enrichment fold change of each CF family within a given cluster relative to the remaining two clusters (Fisher’s exact test, FDR < 0.05; Fig. 3c and Supplementary Data 3).

Fig. 3. CF-family encoding patterns define three putative colonization-associated strategies across gut lineages.

a Calinski–Harabasz index for partitioning around medoids (PAM) clustering of species-level CF-family presence/absence profiles across different values of k. The highest index supported k = 3, corresponding to three CF-defined clusters. b PCoA of species-level CF-family presence/absence profiles based on Jaccard dissimilarity. Points represent microbial species and are colored by PAM cluster assignment. Cluster-1, Cluster-2 and Cluster-3 correspond to metabolism-centric, stress-resistance-oriented and communication-mediated putative colonization strategies, respectively. c Ternary plot showing the relative prevalence of CF families across the three CF-defined clusters. Point color indicates average prevalence. d Distribution of cluster-enriched KOs assigned to the three CF-defined clusters. e KEGG pathway enrichment of cluster-enriched KOs. Pathways are grouped by KEGG BRITE hierarchy. Point size indicates fold enrichment, and color indicates −log₁₀ FDR. f Phylum-level taxonomic composition of species assigned to CF-defined clusters. g Family-level taxonomic distribution across CF-defined clusters for selected prevalent families. For f and g, the overall association between taxonomic composition and CF-defined cluster assignment was assessed using two-sided Fisher’s exact test on the full contingency table, with Cramér’s V used to quantify the strength of association. To identify the taxa contributing to this association, each phylum or family was further tested separately using two-sided Fisher’s exact test by comparing its distribution across the three clusters with that of all remaining taxa. Taxon-specific P values were adjusted for multiple testing using the Benjamini–Hochberg method: * FDR < 0.05; ** FDR < 0.01; *** FDR < 0.001.

Cluster-1-enriched CF families spanned extensive functional diversity, dominated by metabolic activities and transport processes involving central carbon, amino acid, lipid, nucleotide, and vitamin metabolism. Representative CF families included CF37 (pyridoxine kinase), CF33_53 (class-II fructose-bisphosphate aldolase), CF55 (galactokinase), CF70 (LysR-type transcriptional regulator), CF14 (aspartate-ammonia ligase), CF44_51 (outer membrane component protein), CF46 (phosphoribosylformylglycinamidine synthase), and CF22 (thiamine biosynthesis protein ThiF). These characteristics support a putative metabolism-centric colonization strategy among Cluster-1 members, emphasizing efficient substrate utilization and robust metabolic versatility. By contrast, Cluster-2-enriched CF families were functionally associated with redox homeostasis, stress tolerance, and environmental sensing. Representative CF families included CF66 (patatin-like phospholipase), CF43, CF2_7 (ribonucleoside-triphosphate reductase), CF38 (transcriptional regulator HxlR family protein), CF32_36 (chromate transporter) and CF23_58 (haloacid dehalogenase-like hydrolase). This functional configuration suggests a putative stress-resistance-oriented strategy, characterized by oxidative stress mitigation, adaptive regulation and tolerance to potentially harmful environmental conditions. Cluster-3 showed distinct enrichment of CF families associated with carbohydrate uptake, transport and quorum sensing, particularly the phosphotransferase system (PTS). These included CF34 (galactose PTS system EIIC component), CF36 (PTS system EIIBCA component), CF6 (multi-type sugar PTS system EIID component), CF8 (multi-type sugar PTS system EIIC/EIID component), and CF16 (mannose PTS system EIIA component). Additional enriched families included CF41 (S-ribosylhomocysteine lyase), CF63 (DNA polymerase V), CF28_31 (uracil permease), CF68 (ABC transporter substrate-binding protein), CF45 (manganese efflux pump protein) and CF51 (ferrous iron transport protein A). This configuration is consistent with a putative communication-mediated colonization strategy in which carbon-source sensing, nutrient acquisition, transmembrane transport and intercellular signaling may jointly support niche establishment in nutrient-competitive gut environments.

To determine whether CF-defined strategies were linked to broader genomic functional organization, we first assessed the overall concordance between species-level CF-family repertoires and genome-wide KO profiles. A Mantel test between Jaccard dissimilarity matrices calculated from CF-family and KO presence/absence profiles revealed significant concordance between the two functional layers (Mantel r = 0.686, p < 0.001). We next identified KOs enriched in each CF-defined cluster (Fisher’s exact test, FDR < 0.05 and enrichment fold change > 2; Fig. 3d and Supplementary Data 4). KEGG pathway enrichment analysis of these cluster-enriched KOs supported functional diversification across CF-defined strategies (FDR < 0.05; Fig. 3e and Supplementary Data 5). Cluster-1 accounted for the largest proportion of cluster-enriched KOs and was enriched in pathways related to carbon and amino acid metabolism, including butanoate metabolism, propanoate metabolism, carbon fixation pathways in prokaryotes, and arginine and proline metabolism. Cluster-1-enriched KOs were also linked to cell growth, bacterial chemotaxis, and flagellar assembly. Cluster-2-enriched KOs were preferentially associated with redox homeostasis, cell-envelope remodeling, and surface polysaccharide production, including lipopolysaccharide biosynthesis, exopolysaccharide biosynthesis, oxidative phosphorylation, riboflavin and biotin metabolism, as well as ubiquinone and other terpenoid-quinone biosynthesis. By contrast, Cluster-3-enriched KOs were associated with sugar uptake, carbohydrate metabolism and signal transduction or regulatory pathways, including the phosphotransferase system (PTS), galactose metabolism, fructose and mannose metabolism, along with quorum sensing and transcription factors.

Taxonomic mapping revealed strong lineage-associated biases in CF-defined strategy assignment at both the phylum (Cramér’s V = 0.74) and family (Cramér’s V = 0.84) levels (Fisher’s exact test, p < 0.001; Fig. 3f, g). At the phylum level, species within Firmicutes_A were predominantly assigned to the metabolism-centric strategy, whereas Bacteroidota species were mainly associated with the stress-resistance-oriented strategy (Fig. 3f and Supplementary Data 6). By contrast, Actinobacteriota species were more frequently assigned to the communication-mediated strategy. At the family level, 76 families exhibited significant strategy preferences (Fig. 3g and Supplementary Data 7). Species from Lachnospiraceae, Oscillospiraceae and 28 additional families were preferentially associated with the metabolism-centric colonization strategy. Species from Bacteroidaceae, Muribaculaceae and 31 additional families were preferentially associated with the stress-resistance-oriented strategy, whereas species from Coriobacteriaceae, Lactobacillaceae and 11 additional families were more frequently assigned to the communication-mediated strategy. Together, lineage-associated variation in CF-family repertoires and concordant genome-wide functional differentiation are consistent with three putative colonization strategies in the human gut.

Applying the CF framework to IBD reveals recurrent remodeling of colonization-associated repertoires

We then applied the CF framework to IBD, in which microbial dysbiosis is well documented but colonization-associated mechanisms remain unclear. We curated 2858 publicly available human fecal metagenomic samples from 10 independent inflammatory bowel disease (IBD) cohorts, encompassing multiple disease subtypes and research centers (Fig. 4a). Using the 78,579 CF homologs identified above as reference sequences, we quantified the relative abundance of CF homologs across all metagenomic samples by read mapping.

Fig. 4. Applying the CF framework to IBD reveals recurrent remodeling of colonization-associated repertoires.

a Sample sizes for the 19 case–control comparisons across ten IBD metagenomic cohorts. Bars indicate the numbers of case and control samples in each comparison. CD, Crohn’s disease; UC, ulcerative colitis; IC, indeterminate colitis; IBD, inflammatory bowel disease. b Relative abundance profiles of dominant CF families across the 19 comparisons. Stacked bars show the mean relative abundance of dominant CF families in each comparison. c, d Disease-associated alterations in CF-gene alpha-diversity across comparisons, including Shannon diversity (c) and richness (d). Values indicate log₂ fold changes between case and control group means. Statistical significance was assessed using two-sided Wilcoxon rank-sum tests followed by Benjamini–Hochberg correction. Asterisks in c and d indicate significance: * FDR < 0.05; ** FDR < 0.01; *** FDR < 0.001. e PERMANOVA effect sizes showing the fraction of CF-gene compositional variation explained by disease status in each comparison. Bars indicate adjusted R² values. P values were calculated using a one-sided upper-tail permutation test based on the pseudo-F statistic with 1000 permutations. Asterisks indicate significance: * p < 0.05; ** p < 0.01; *** p < 0.001. f–h Matched randomization analyses comparing average gene-level statistics from CF-like POG-associated genes with matched random POG-derived gene backgrounds. Density curves represent empirical null distributions generated from genes assigned to matched random POG sets, and red lines indicate the observed average CF-like POG-associated gene statistics across comparisons for Shannon diversity (f), richness (g), and compositional effect size (h). SES values indicate standardized deviations from the matched random background. For alpha-diversity statistics, positive SES values indicate weaker case-associated depletion than matched random backgrounds. Empirical one-sided P values were calculated as the proportion of matched random POG sets yielding statistics equal to or greater than the observed statistic.

Across the 19 comparisons, CF-family abundance profiles were dominated by a small set of recurrent CF families. CF5 and CF0_0 were the most abundant families, with mean relative abundances within CF-family profiles of 11.1 ± 1.1% and 10.9 ± 1.7%, respectively, followed by CF31 (formate C-acetyltransferase; 3.6 ± 0.7%), CF4 (3.3 ± 0.6%) and CF35 (acetate kinase; 3.2 ± 0.5%; Fig. 4b). The relative abundance hierarchy of these dominant CF families was broadly consistent across cohorts and disease subtypes, suggesting a broadly consistent abundance structure among dominant CF families in the human gut microbiome.

We next characterized IBD-associated alterations in CF gene-level diversity. Across the 19 comparisons, 14 (73.7%) exhibited lower CF-gene Shannon diversity in cases than in controls, with 8 reaching statistical significance (Wilcoxon rank-sum test, FDR < 0.05; Fig. 4c). CF-gene richness showed an even more consistent trend, with 16 comparisons (84.2%) showing decreased richness in cases and 12 reaching statistical significance (Wilcoxon rank-sum test, FDR < 0.05; Fig. 4d). Thus, IBD-associated dysbiosis was accompanied by recurrent reductions in CF-gene diversity, although the magnitude and statistical strength of these alterations varied across cohorts and disease subtypes. PERMANOVA further indicated that disease status was associated with detectable shifts in overall CF-gene composition in each pairwise comparison. Disease status explained a generally modest fraction of CF-gene compositional variation, with effect size estimates averaging 4.3 ± 6.3% (range, 0.4–28.3%) across comparisons (Fig. 4e).

To contextualize the magnitude of these CF-gene-associated changes, we compared the observed diversity and compositional effects with gene backgrounds derived from matched random POG sets. For Shannon diversity, the observed coefficient was positioned toward the attenuated side of the empirical null distribution, indicating weaker case-associated depletion than that observed for matched random backgrounds (SES = 2.028, p = 0.013; Fig. 4f). A similar but non-significant tendency was observed for richness (SES = 1.082, p = 0.149; Fig. 4g), whereas the proportion of variation in CF-like POG gene composition explained by disease status was broadly comparable to the random expectation (SES = −0.287, p = 0.581; Fig. 4h). These findings indicate reproducible but relatively restrained remodeling of colonization-associated repertoires in IBD.

IBD-associated CF remodeling maps to specific colonization-associated functions and carrier lineages

Having observed recurrent CF repertoire remodeling in IBD-associated dysbiosis, we next asked whether these alterations could be linked to species-level community shifts and resolved into specific colonization-associated functions and microbial carrier lineages. Across the ten cohorts, Procrustes analysis revealed concordance between CF-family abundance profiles and species-level taxonomic abundance profiles (mean Procrustes m² = 0.388 ± 0.139, permutation test: p < 0.001; Fig. 5a and Supplementary Fig. 3). However, the magnitude of this concordance was comparable to that observed for matched random backgrounds (SES = 0.721, p = 0.198; Supplementary Fig. 4a). Reciprocal explanatory analyses further showed that species composition explained a larger fraction of CF-family profile variation than CF-family profiles explained species-level community variation (adjusted R² = 0.878 ± 0.038 versus 0.547 ± 0.05; Fig. 5b, c). Nevertheless, neither directional relationship was stronger than the matched background expectation at the CF-like POG level (Supplementary Fig. 4b, c).

Fig. 5. IBD-associated CF remodeling maps to specific colonization-associated functions and carrier lineages.

a Procrustes analysis measuring concordance between CF-family abundance profiles and species-level taxonomic profiles across ten cohorts. Statistical significance was assessed using a one-sided upper-tail Procrustes permutation test with 1000 permutations: *** p = 0.001. b, c Reciprocal explanatory analyses showing the proportion of species-level community variation explained by CF-family abundance profiles (b) and the proportion of CF-family profile variation explained by species-level community composition (c). Effect sizes are shown as adjusted R². Explanatory variables were derived from principal components cumulatively explaining 95% of the variance. d Heatmap showing the explanatory capacity of individual CF families for species-level community structure across cohorts, measured as adjusted R². The bar plot above the panel represents the average adjusted R² for each CF family. Color bars below the x-axis labels denote the CF-defined strategies. For b–d, P values were calculated using one-sided upper-tail permutation tests based on the pseudo-F statistic with 1000 permutations: * p < 0.05; ** p < 0.01; ***/+ p = 0.001. e Gene set enrichment analysis (GSEA)-like carrier enrichment analysis testing whether species carrying each CF family were biased toward IBD- or healthy control (HC)-associated taxa. Positive normalized enrichment score (NES) values indicate IBD-biased CF-family carrier sets, whereas negative NES values indicate HC-biased CF-family carrier sets. Points are colored by putative colonization strategy. f Family-level taxonomic composition of leading-edge carrier species from significant CF-family carrier enrichment results. g Network visualization linking representative genera, leading-edge carrier species, and CF families from significant CF-family carrier sets. Circular nodes denote CF families, and diamond nodes denote species. Node border color indicates whether each node is IBD- or HC-biased based on carrier enrichment for CF-family nodes and species-level meta-analysis estimates for species nodes. CF-family nodes are colored by putative colonization strategy, and species nodes are colored by genus-level taxonomy.

We next resolved the CF–species relationship at the level of individual CF families to identify CF families most strongly associated with species-level community variation. Explanatory strength varied markedly across CF families, indicating that the CF–species linkage was not evenly distributed across the repertoire (Fig. 5d). CF families with high explanatory values were detected across all three putative colonization-associated programs, but their contributions were uneven. In particular, PTS-related CF families assigned to the communication-mediated strategy, including CF8, CF16, CF34, CF6, and CF36, were among the highest-ranking families, with average adjusted R² values ranging from 0.092 to 0.103. Stress-resistance-oriented families, including CF32_36, CF23_58 and CF66, also showed relatively high explanatory values. Meanwhile, MATE-related CF0_0, CF0_6 and CF5 had notable explanatory capacity, with average adjusted R² values ranging from 0.075 to 0.099. By comparison, metabolism-centric families generally showed more moderate explanatory values than these high-ranking communication-mediated and stress-resistance-oriented families, although explanatory strength varied within this strategy. At the lower end of the distribution, CF64 (YitT-family uncharacterized membrane protein) and CF56 (U32 family peptidase) had limited associations with community composition. These results suggest that the explanatory strength varied substantially across CF families rather than being distributed uniformly across the CF repertoire.

To determine the microbial carrier lineages underlying IBD-associated CF remodeling, we first compared the encoded CF repertoires of species consistently associated with IBD or healthy controls (HC) across cohorts. IBD-enriched and HC-enriched species were identified by meta-analysis across the 19 pairwise case–control comparisons (FDR < 0.05; Supplementary Fig. 4d and Supplementary Data 8) and integrated with species-level CF-family presence/absence profiles. Total CF-family richness did not differ significantly between IBD-enriched and HC-enriched species (Supplementary Fig. 4e). However, strategy-specific analyses revealed significantly lower metabolism-centric and stress-resistance-oriented CF-family richness in IBD-enriched species than in HC-enriched species, whereas communication-mediated CF richness showed only a marginal difference (communication-mediated, p = 0.056; metabolism-centric, p = 0.0004; stress-resistance-oriented, p = 2.6 × 10⁻¹¹; Supplementary Fig. 4f–h). IBD-enriched species also had higher within-group pairwise Jaccard dissimilarities than did HC-enriched species across total CF-family profiles and each strategy-specific CF subset, with the largest differences observed for communication-mediated and metabolism-centric CF families (Supplementary Fig. 4i–l). These results indicate that IBD-associated species are characterized by altered CF repertoire organization, reflected in strategy-specific richness differences and greater compositional heterogeneity of encoded CF-family profiles.

To determine whether specific CF families were preferentially encoded by IBD- or HC-associated taxa, we performed a gene set enrichment analysis (GSEA)-like carrier enrichment analysis. This analysis revealed an asymmetry in CF carrier breadth: CF families with broad carrier species sets were predominantly enriched toward the HC-biased end of the species ranking, whereas several IBD-biased CF families were supported by narrower carrier species sets (FDR < 0.05; 33 HC-biased versus 14 IBD-biased CF carrier sets; Fig. 5e and Supplementary Data 9). We then identified the key carrier species contributing to each significant CF-family carrier enrichment and summarized their taxonomic composition (Fig. 5f). IBD-biased CF-family carrier sets were frequently supported by species from Streptococcaceae and Enterobacteriaceae, whereas HC-biased carrier sets were supported by species from commensal-associated families, including Ruminococcaceae, Lachnospiraceae, and Oscillospiraceae. Network visualization revealed a modular organization of IBD- and HC-biased CF-family carrier sets (Fig. 5g). IBD-biased carrier modules were mainly linked to inflammation-associated genera such as Fusobacterium, Streptococcus and Enterococcus, whereas HC-biased modules were dominated by commensal-associated genera such as Roseburia and Faecalibacterium. Notably, HC-biased modules contained multiple metabolism-centric CF families and HC-associated carrier species, including the HC-associated carrier species Bifidobacterium adolescentis and Bifidobacterium ruminantium, whereas Bifidobacterium breve, Bifidobacterium bifidum and Bifidobacterium infantis contributed to IBD-biased carrier patterns. Together, these results suggest that IBD-associated CF remodeling is organized by distinct carrier-lineage architectures, with HC-biased CF families supported by broader commensal-associated carrier sets and IBD-biased CF families linked to narrower, inflammation-associated carrier lineages.

Feature compression identifies a compact and biologically interpretable 13-CF-family signature of IBD status

Building on the recurrent CF remodeling associated with IBD, we next sought to identify CF-family signatures that were reproducibly associated with IBD across cohorts. DESeq2 analysis showed that nearly all CF families, except CF13 (xanthine phosphoribosyltransferase), were significantly altered in at least one case–control comparison, with either enrichment or depletion in patients with IBD (FDR < 0.05; Fig. 6a). The number and strength of differential CF families varied across cohorts and disease subtypes, with more widespread CF alterations observed in Crohn’s disease (CD) than in ulcerative colitis (UC). To distinguish recurrent cross-cohort signals from cohort-specific alterations, we performed a meta-analysis across the 19 pairwise case–control comparisons, which identified a 46-CF-family signature significantly associated with IBD status, including 30 IBD-enriched and 16 HC-enriched CF families (FDR < 0.05; Fig. 6b and Supplementary Data 10). IBD-enriched CF families included multiple PTS-related families, such as CF36, CF8, CF16, CF6 and CF34, suggesting increased representation of sugar transport and carbohydrate-utilization-associated colonization functions. Other IBD-enriched CF families were associated with transport, hydrolysis and protein degradation, including CF10 (branched-chain amino acid transporter, LIV-CS family protein), CF23_25 (sucrose-6F-phosphate phosphohydrolase), CF23_58, CF26 (aminopeptidase C), and CF11. In contrast, HC-enriched families included CF20, CF30 (flavorubredoxin/flavodoxin-like redox protein) and CF1, which were associated with iron-sulfur cluster assembly²², flavodoxin-like redox reactions, and radical SAM-dependent anaerobic nucleotide metabolism, respectively. Additional HC-enriched CF families, including CF48_61, CF4, and CF46, were involved in nucleotide biosynthesis and metabolism. Together, these results define a recurrent disease-associated CF-family signature that reflects functional shifts in the IBD gut microbiome.

Fig. 6. Feature compression identifies a compact and biologically interpretable 13-CF-family signature of IBD status.

a Heatmap of differential CF-family abundance across 19 pairwise case–control comparisons. Colors indicate log₂-transformed fold changes. Significance was assessed using DESeq2 with Benjamini–Hochberg (BH) adjustment: * FDR < 0.05; ** FDR < 0.01; + FDR < 0.001. b Meta-analysis of CF-family associations with IBD status across 19 comparisons. Points indicate pooled Hedges’ g estimates from random-effects meta-analyses, and error bars indicate the corresponding 95% confidence intervals. Positive and negative estimates indicate enrichment in IBD and HC, respectively; yellow and blue denote these two directions, respectively. The dashed line indicates no effect (Hedges’ g = 0). Two-sided P values from the random-effects models were adjusted using the BH method: * FDR < 0.05; ** FDR < 0.01; *** FDR < 0.001. c, d Receiver operating characteristic (ROC) curves from 10-fold cross-validated random forest classifiers using all 79 CF families (c) and 46 meta-analysis-derived CF families (d) to discriminate IBD patients from HC. e Inter-feature redundancy of all CF-like POGs and matched random POG sets across cohorts. Redundancy was quantified within each cohort as the mean absolute pairwise Spearman correlation among features. Empirical one-sided P values were calculated as the proportion of matched random POG sets yielding statistics equal to or greater than the observed statistic. f Feature-compression analysis of CF-like POGs. CF-like POGs were ranked by average random forest variable importance across cohort-specific models, and average area under the receiver operating characteristic curve (AUC) was evaluated using increasing numbers of top-ranked CF-like POGs. The dashed line indicates the selected compact 15-CF-like POG panel. g Matched randomization analysis of the selected 15-CF-like-POG panel against size-matched random POG panels. The red line indicates the observed AUC, and the density curve represents the empirical null distribution. Empirical one-sided P values were calculated as the proportion of matched random POG sets yielding statistics equal to or greater than the observed statistic. h ROC curves from 10-fold cross-validated random forest classifiers using the compact 13 CF-family panel to discriminate IBD patients from HC.

We next evaluated the discriminatory information captured by these CF-family profiles. Random forest classifiers based on all 79 CF families and the 46 meta-analysis-derived CF families achieved similar average area under the receiver operating characteristic curve (AUC) values of 0.838 ± 0.099 and 0.833 ± 0.092, respectively (Fig. 6c, d and Supplementary Fig. 5a), indicating that the recurrent 46-CF-family signature retained most of the discriminatory information captured by the full CF repertoire. However, in a same-framework sensitivity analysis, classifiers based on all CF-like POGs did not outperform matched random protein-family backgrounds with comparable feature numbers (SES = −2.266, p = 0.985; Supplementary Fig. 5b). Together with the higher inter-feature redundancy observed among CF-like POGs than among matched random POG sets (Fig. 6e), this result indicated that the discriminatory information in the full CF-like POG set was concentrated in correlated, partially redundant features rather than broadly distributed across the repertoire.

We therefore performed feature-compression analyses to determine whether IBD-associated CF information could be retained in a smaller and more biologically interpretable feature set. At the CF-like POG level, model performance rapidly reached a plateau, and a selected 15-feature CF-like POG panel achieved an AUC of 0.826 and significantly outperformed random 15-POG sets (SES = 2.014, p = 0.020; Fig. 6f, g). Applying the same feature-compression strategy to curated CF-family profiles identified a compact 13-CF-family panel, which achieved an average AUC of 0.824 ± 0.094 across cohorts, only slightly lower than those of the 79- and 46-CF-family models despite using substantially fewer features (Fig. 6h, Supplementary Fig. 5a, c and Supplementary Data 11). These results indicate that much of the discriminatory information captured by the larger CF-family panels could be compressed into a small set of coordinated and biologically interpretable functional features.

For benchmarking, classifiers based on high-dimensional species- and genus-level profiles achieved higher average AUCs of 0.905 ± 0.100 and 0.898 ± 0.115, respectively (Supplementary Fig. 5a, d and e). Size-matched random taxonomic analyses showed that the performance of the compact 13-CF-family signature was positively shifted relative to random species- and genus-level feature sets, although the comparisons did not reach statistical significance (species level, SES = 1.206, p = 0.102; genus level, SES = 1.165, p = 0.118; Supplementary Fig. 5f–k). By contrast, the 79- and 46-CF-family models showed no clear advantage over their corresponding random taxonomic backgrounds. In a complementary analysis based on random forest feature importance, top-ranked CF-family panels generally underperformed the corresponding top-ranked species- and genus-level panels (Supplementary Fig. 5l). Notably, the performance gap was smallest in the 13-feature comparison, further supporting the compact 13-CF-family signature as the most feature-efficient CF-based representation of IBD status. Cross-dataset validation showed comparable generalizability across the 79-, 46-, and 13-CF-family models, indicating that feature compression did not markedly compromise transfer performance, with AUC values exceeding 0.75 in several test cohorts (Supplementary Fig. 5m–o). Thus, although not intended to replace high-dimensional taxonomic models, the 13-CF-family signature provided a compact and biologically interpretable representation of IBD-associated dysbiosis by projecting disease-associated community shifts onto colonization-related functions.

Metatranscriptomic profiling extends the CF atlas to transcriptional activity in IBD

To determine how the genomically encoded CF repertoire is deployed transcriptionally in the human gut and whether its expression retains disease-associated signals, we analyzed CF-family expression profiles in two publicly available IBD metatranscriptomic cohorts with matched metagenomic data. Across both datasets, CF3 (rubrerythrin), CF43, and CF4 were among the most highly expressed CF families. The most highly expressed CF families were broadly concordant between datasets, suggesting a shared in vivo transcriptional hierarchy of CF families in the human gut microbiome (Fig. 7a, b). Matched metagenomic and metatranscriptomic samples showed strong positive correlations between average CF-family abundance and expression levels (Spearman r = 0.95 and 0.90 in LloydPriceJ_2019 and SchirmerM_2018, respectively; Fig. 7c, d). However, CF transcriptional activity was not simply proportional to metagenomic abundance. Relative to the one-to-one abundance–expression relationship, highly abundant families such as CF0_0 and CF5 showed lower transcriptional output than expected, whereas CF3, CF43 and CF4 were transcriptionally overrepresented. Matched POG randomization indicated that CF-family abundance–expression coupling was strong but did not significantly exceed that observed for comparable backgrounds (LloydPriceJ_2019, SES = 0.94, p = 0.166; SchirmerM_2018, SES = 1.338, p = 0.067; Supplementary Fig. 6a, b). These results indicate that CF-family transcriptional activity exhibits a shared expression hierarchy and family-specific transcriptional biases across cohorts.

Fig. 7. Metatranscriptomic profiling extends the CF atlas to transcriptional activity in IBD.

a, b Ridge plots illustrating the expression-level distributions of the 20 most highly expressed CF families in LloydPriceJ_2019 (a) and SchirmerM_2018 (b). Expression levels are shown as log₁₀-transformed TPM values. c, d Relationships between average CF-family metagenomic abundance and metatranscriptomic expression levels in matched samples from LloydPriceJ_2019 (c) and SchirmerM_2018 (d). Each point represents one CF family. The n value shown above each panel denotes the number of matched metagenomic–metatranscriptomic samples used to calculate the mean abundance and expression level of each CF family. The blue line indicates the fitted ordinary least-squares linear regression and the black line indicates the one-to-one relationship. Spearman’s rank correlation coefficients and linear regression statistics are shown. Two-sided P values for Spearman’s correlations tested the null hypothesis that Spearman’s r was zero, whereas two-sided P values for the linear regressions tested the null hypothesis that the regression slope was zero. e, f Metatranscriptomic mapping rates to CF genes across HC, CD and UC samples in LloydPriceJ_2019 (e) and SchirmerM_2018 (f). g–j Diversity of expressed CF genes across disease groups, including richness in LloydPriceJ_2019 (g) and SchirmerM_2018 (h), and Shannon diversity in LloydPriceJ_2019 (i) and SchirmerM_2018 (j). For e–j, exact n values are shown in the corresponding panels; n denotes the number of metatranscriptomic samples included in each group. Each point represents one metatranscriptomic sample. Boxes represent the interquartile range (IQR) from the first quartile to the third quartile, with the internal line indicating the median. Whiskers extend to the minimum and maximum values within 1.5 × IQR of the box boundaries. Points beyond the whiskers represent outliers. P values were calculated using two-sided Wilcoxon rank-sum tests. k ROC curves of random forest classifiers based on CF-family expression profiles for discriminating IBD patients from HC in LloydPriceJ_2019 and SchirmerM_2018. AUC values and 95% confidence intervals are shown. l Matched randomization analysis of the expression panel comprising the top-ranked 18 CF-like POG features in LloydPriceJ_2019. The red line indicates the observed AUC, and the density curve represents the empirical null distribution from size-matched random expressed POG panels. Empirical one-sided P values were calculated as the proportion of matched random POG sets yielding statistics equal to or greater than the observed statistic. m Feature-compression analysis of CF-family expression profiles in LloydPriceJ_2019. AUC values were evaluated using increasing numbers of top-ranked CF families, and the dashed line indicates the selected compact 21-CF-family expression panel. n Top-ranked CF-family expression features contributing to expression-based disease discrimination in LloydPriceJ_2019, ordered by mean decrease in accuracy (MDA) from random forest models.

We next evaluated whether expressed CF-gene diversity was associated with IBD. Overall metatranscriptomic mapping rates were comparable between IBD and HC samples (Fig. 7e, f). Expressed CF-gene Shannon diversity and richness were not consistently altered across datasets or disease subtypes (Fig. 7g–j). The main exception was expressed CF-gene richness in LloydPriceJ_2019, which was significantly reduced in patients with CD compared with both HC and UC samples (Wilcoxon rank-sum test, p < 0.01), whereas no significant richness differences were observed in SchirmerM_2018. Thus, expressed CF-gene diversity was largely stable, with cohort- and subtype-dependent richness alterations rather than a consistent IBD-wide loss of transcriptional diversity.

Nevertheless, CF-family expression profiles retained disease-discriminatory information. Random forest classifiers based on CF-family expression profiles achieved strong discrimination in LloydPriceJ_2019 (AUC = 0.879), comparable to the corresponding metagenomic abundance model, but showed weaker performance in SchirmerM_2018 (AUC = 0.678; Fig. 7k). Randomization analysis showed that full CF-like POG expression models did not consistently outperform matched random expressed POG backgrounds (LloydPriceJ_2019, SES = −2.536, p = 0.990; SchirmerM_2018, SES = −0.112, p = 0.543; Supplementary Fig. 6c, d). This lack of a full-feature advantage was accompanied by pronounced inter-feature redundancy among expressed CF-like POGs, particularly in LloydPriceJ_2019 (SES = 13.031, p = 0.001; Supplementary Fig. 6e).

Given the stronger expression-level disease signal in LloydPriceJ_2019, we tested whether this signal could be compressed into a smaller set of CF-associated expression features. At the CF-like POG level, the top-ranked 18 expression features achieved an average AUC of 0.842 and outperformed size-matched random expressed POG backgrounds, although the comparison did not reach conventional statistical significance (SES = 1.241, p = 0.094; Fig. 7l and Supplementary Fig. 6f). Applying the same ranking strategy to curated CF-family expression profiles identified a compact 21-CF-family expression panel comprising the top-ranked features in LloydPriceJ_2019, with an average AUC of 0.872 (Fig. 7m). The most important features included CF23_58, CF37, CF5, CF30 and CF11, highlighting hydrolysis, phosphorylation-related activity, MATE efflux, flavodoxin-related redox activity and acetyltransferase-related functions as major contributors to expression-based discrimination (Fig. 7n). Taken together, these analyses reveal a shared transcriptional hierarchy across the encoded CF repertoire, alongside cohort-dependent disease signals concentrated in a limited subset of expressed CF families. By adding information on functional activity to metagenomic abundance profiles, CF expression profiling provides a complementary view of which colonization-associated functions are actively deployed in IBD-associated gut communities.

Discussion

CF research has historically focused on specific pathogens or model commensals^{15–17,23,24}. Although high-throughput sequencing and advanced computational frameworks have enabled large-scale discovery of gut microbial colonization-associated genes^18,25,26, the organization of these factors across the human gut microbiome and their relevance to disease-associated community remodeling have remained incompletely characterized. In this study, we constructed an integrative catalog of 7,056,385 CF homologs across 289,022 CF-encoding genomes from 4716 microbial species. This atlas extends CF research beyond selected organisms and reference CF families to a microbiome-wide view of genome-encoded colonization potential. It revealed the widespread distribution, phylogenetically structured organization, and functional stratification of CF repertoires, and connected these patterns to IBD-associated dysbiosis, microbial carrier taxa, and transcriptional activity. By organizing colonization-associated genes into lineage-linked functional strategies and applying this framework to multi-cohort IBD datasets, our study provides a mechanistically informed framework for interpreting disease-associated microbiome remodeling from the perspective of colonization function, complementing taxonomic descriptions of gut microbiome dysbiosis.

The human gut CF atlas revealed a near-universal but lineage-differentiated CF landscape. Most gut microbial genomes (99.9%; 289,022/289,231) encoded at least one CF gene, indicating that CF-encoding potential is widespread among human gut microbes. However, CF-family repertoire breadth varied substantially among taxa, with distinct lineages possessing characteristic CF repertoires, potentially reflecting adaptation to differences in nutrient availability, host-derived stressors, mucosal habitats and microbial competition^7,8,13,14,27. Notably, the weaker taxonomic structuring of CF-like POG profiles relative to matched backgrounds suggests that lineage differentiation alone does not fully explain CF repertoire organization. CF-family profiles should therefore not be viewed as taxonomy-independent features or as evidence of uniquely strong phylogenetic structuring, but rather as a broadly shared yet lineage-differentiated functional layer of colonization-associated potential. This organization may reflect the joint influence of shared colonization demands across diverse gut microbes and lineage-specific ecological specialization.

This lineage-differentiated organization provides a basis for interpreting the three CF-defined strategies identified in this study. Colonization by gut microbes proceeds through defined spatiotemporal stages, including gastrointestinal transit, niche adaptation, population expansion and long-term persistence^7,8,27. Successfully navigating microenvironmental challenges, such as anaerobiosis, pH fluctuations, digestive enzymes, mucus barriers and peristalsis, requires distinct colonization-associated functional capacities^3,6,27. In this context, the putative metabolism-centric, stress-resistance-oriented and communication-mediated CF strategies may represent alternative colonization-associated solutions encoded by different gut microbial lineages. The metabolism-centric strategy is consistent with the importance of substrate utilization and metabolic niche occupation during engraftment^14,28. The stress-resistance-oriented strategy highlights redox homeostasis, cell-envelope remodeling and tolerance to host-derived stress²⁹, whereas the communication-mediated strategy emphasizes PTS-dependent nutrient sensing, quorum sensing and regulatory functions^30–32. Genome-wide KO analyses supported these interpretations, indicating that CF-defined strategies reflect broader functional programs rather than isolated marker families. This framework therefore provides a functional basis for understanding how gut microbial lineages differ in their encoded colonization potential.

The close association between microbiota dysbiosis and human disease motivates investigation of CF dynamics under pathological conditions^19,33–35. In IBD, CF-gene abundance profiles captured recurrent but relatively constrained signatures of dysbiosis across cohorts. This constrained pattern suggests that CF repertoires do not appear globally destabilized in IBD, but instead form a conserved and buffered colonization-associated layer that undergoes selective remodeling in the context of inflammatory and ecological disturbance. Metatranscriptomic analyses extended this interpretation to the level of transcriptional activity: CF-family expression was strongly coupled to metagenomic abundance and exhibited a shared transcriptional hierarchy across cohorts, whereas disease-associated expression signals were cohort-dependent and concentrated in specific expressed CF-family features rather than broad transcriptional diversity loss. Together, these abundance- and expression-level patterns suggest that CF repertoires represent a relatively stable colonization-associated functional layer, within which IBD-associated dysbiosis is reflected by selective remodeling of specific CF families rather than a global disruption of genome-encoded colonization potential.

PTS-related CF families provide a notable example of this selective remodeling, as they were repeatedly highlighted in both family-level explanatory analyses and cross-cohort IBD-associated signatures. Beyond sugar uptake, the bacterial PTS functions as a regulatory hub coordinating carbon utilization, metabolic reprogramming, biofilm formation and virulence^36,37. Experimental evidence supports its relevance to inflammatory colonization niches: the PTS in Enterococcus faecalis mediates glucosamine uptake and exacerbates experimental intestinal inflammation³⁸, and within-host evolution of Enterococcus gallinarum can select for mutations in PTS genes, such as manX and manY, that enhance mucosal colonization and promote systemic translocation, driving extra-intestinal inflammation³⁹. Together with the recurrent enrichment of PTS-related CF families in IBD-associated signatures across multiple cohorts, these observations suggest that carbohydrate utilization and nutrient-sensing programs may contribute to bacterial fitness in inflamed gut environments^21,40. Nevertheless, the present data remain associative, and functional experiments are needed to determine whether specific PTS-related CF genes causally contribute to IBD-associated colonization fitness or inflammatory progression.

A key challenge in interpreting CF-family remodeling is that CF-family abundance profiles aggregate homologous functions across microbial taxa. As a result, the same CF-family signal may have different ecological implications depending on which microbial lineages encode it. The carrier-resolved analyses addressed this issue by linking disease-associated CF-family shifts back to their microbial carriers. By identifying which health-associated or inflammation-associated lineages encoded specific CF families, this analysis placed CF remodeling within a microbial ecological context rather than treating it as an abstract functional change. Thus, carrier-level resolution provides a bridge between taxonomic dysbiosis and colonization-associated functional interpretation.

The classification analyses clarify the value and limits of CF-family abundance- and expression-based signatures as predictive features. At the metagenomic abundance level, CF-family models showed lower discriminatory performance than species- and genus-level taxonomic profiles, which is biologically expected because taxonomic profiles retain broader community-compositional and potentially strain-associated information^41–43. By contrast, CF-family profiles aggregate homologous genes from multiple microbial lineages into a smaller set of colonization-associated functional features. This aggregation improves biological interpretability but compresses part of the taxonomic and strain-level information that may contribute to classification. The value of CF-family profiles therefore lies in providing a compact and functionally informed representation of disease-associated microbial dysbiosis. This positioning is consistent with recent efforts to translate microbiome signatures into more targeted and interpretable diagnostic frameworks^44,45.

However, this study has several limitations. First, current reference genome catalogs remain incomplete, and CF repertoires will likely expand as additional gut microbial genomes and strain-resolved assemblies become available. Second, although matched random POG analyses provided empirical background controls within the same analytical framework, CF-like POGs and curated CF families are not identical analytical units; therefore, randomization results should be interpreted as sensitivity analyses rather than direct surrogates for the 79 curated CF families. Third, carrier-resolved analyses depend on reference genome representation and species-level CF presence criteria, which may underrepresent strain-level heterogeneity. Fourth, metatranscriptomic analyses were limited to two cohorts and showed cohort dependence. Finally, all disease associations in this study remain observational, and the causal roles of CFs in disease-associated dysbiosis require functional genomics and animal-model validation.

In summary, this study establishes an integrative atlas of CF repertoires across the human gut microbiome and shows that these repertoires form a broadly distributed, lineage-associated and functionally interpretable layer of microbial colonization potential. By linking CF diversity, phylogenetic organization, colonization-associated strategies, disease-associated carrier taxa and transcriptional activity, our work suggests that IBD-associated dysbiosis involves selective remodeling of specific colonization-associated functions and their microbial carriers rather than wholesale disruption of the CF repertoire. More broadly, CF-centered studies may help prioritize colonization-associated functions and carrier taxa for mechanistic validation and future biomarker- and intervention-oriented studies^45,46.

Methods

Data curation

A total of 27,096 reference CF protein sequences, representing 79 curated CF families, were obtained from the genotype–habitat association project (GHA). The UHGG and UHGP-90 catalogs⁴⁷ (v2.0.2), comprising 289,231 nonredundant genomes and more than 13.8 million nonredundant proteins clustered at 90% amino acid identity from the human gut microbiome, were downloaded from the EBI MGnify database.

A PubMed-based literature search was conducted to identify human gut metagenomic studies related to inflammatory bowel disease (IBD), Crohn’s disease (CD), and ulcerative colitis (UC). In total, 10 IBD studies were included, comprising 10 shotgun metagenomic datasets and 2 metatranscriptomic datasets with available metadata. After quality control and metadata harmonization, 2,858 metagenomic samples and 808 metatranscriptomic samples were retained for downstream analyses. The metagenomic datasets yielded 19 case–control comparisons. Raw sequencing data were retrieved from the EBI ENA database under the accession numbers listed in Supplementary Data 12.

Identification of CF homologs in the human gut microbiome

UHGP-90 protein sequences were searched against the GHA CF reference proteins using DIAMOND⁴⁸ blastp (v2.1.11.165). For each UHGP-90 query protein, the highest-scoring reference CF hit was retained based on its bit score. Proteins were annotated as CF homologs when they satisfied all of the following criteria: e-value < 1 × 10⁻¹⁰, amino acid identity > 50%, query coverage > 70% and subject coverage > 70%. Functional annotations of reference CF proteins and UHGP-90 proteins were generated using eggNOG-mapper⁴⁹ (v2.1, database v5.0), from which KEGG Orthology (KO) and Pfam annotations were extracted.

Based on UHGP-90 cluster membership information, CF homologs identified among representative genes were propagated to the corresponding member genes and mapped to the UHGG genome catalog. A CF gene was considered present in a species if it was detected in at least one genome assigned to that species. A CF family was considered present in a species if homologs assigned to that CF family were detected in at least half of the genomes belonging to that species. Species-level phylogenetic trees were constructed using PhyloPhlAn⁵⁰ (v3.1.68) and visualized using iTOL (v7.3, https://itol.embl.de/).

Metagenomic and metatranscriptomic read processing

Raw sequencing reads were quality-filtered using fastp⁵¹ (v0.24). Adapter sequences were trimmed, and reads were discarded if they contained more than 55 bases with quality scores <20, more than five ambiguous bases, or low-complexity sequences. Human-derived reads were removed by mapping against the human reference genome GRCh38 (GCF_000001405.40) using Bowtie2⁵² (v2.5.4) with default parameters. To quantify CF-gene abundance and expression, high-quality metagenomic and metatranscriptomic reads were mapped to the nucleotide sequences of CF homologs using Bowtie2. Alignment statistics and coverage were extracted using SAMtools (v1.21), and gene-level abundance and expression values were calculated as TPM-normalized values using custom scripts⁵³. Briefly, read counts for each CF gene were divided by gene length, scaled by the sum of length-normalized read counts across all CF genes in each sample, and multiplied by 10⁶, thereby normalizing for gene length and sequencing depth. A CF gene was considered detected in a metagenomic sample if mapped reads covered at least 50% of its nucleotide length, and was considered expressed in a metatranscriptomic sample if mapped reads covered at least 5% of its nucleotide length. Unless otherwise specified, sample-level CF abundance and expression analyses were performed at the CF-family level using TPM profiles generated by summing TPM values of genes assigned to the same CF family. For DESeq2-based differential abundance analysis, raw read counts were aggregated at the CF-family level and used as the input count matrix.

Taxonomic profiling of gut metagenomes

The UHGG database was used as the reference for taxonomic profiling of fecal metagenomes. Taxonomic classification of quality-controlled reads was performed using Kraken2⁵⁴ (v2.1.3). Taxonomic abundance profiles at multiple ranks, from phylum to species, were re-estimated using Bracken (v3.0.1).

Construction of POG backgrounds and matched random protein-family sets

To construct a protein-family background within the same analytical framework¹⁸, UHGP-90 proteins corresponding to genes detected in metagenomic samples with a sample prevalence > 1% were retained. Sequence similarities among the selected proteins were identified using an all-against-all DIAMOND blastp search with an e-value cutoff of 1 × 10⁻⁵. A pairwise similarity network was then constructed from the alignments, in which each protein represented a node and each significant hit defined an edge. Following the MCL workflow⁵⁵ (v22-282), e-values were transformed into edge similarity weights using mcxload with the “--stream-neg-log10 -stream-tf ceil(200)” options. The resulting weighted graph was clustered using MCL with an inflation parameter of 2.5, and the identified clusters were defined as protein operational groups (POGs).

Because the original CF catalog was defined at the curated reference-family level, whereas POGs represent sequence-resolved operational protein families, CF-like POGs were used as a same-framework sensitivity control rather than direct substitutes for the 79 curated CF families. Using the CF homolog annotations generated above, UHGP-90 proteins were assigned to their reconstructed POGs based on cluster membership information. For each POG, we recorded the number of CF-hit proteins, total family size, CF-hit fraction within the POG, and the dominant reference CF label. POGs containing at least three proteins and a CF-hit fraction ≥ 0.1 were defined as CF-like POGs, yielding 73 CF-like POGs for subsequent randomization analyses.

Matched random POG sets were generated as empirical null backgrounds for CF-like POGs. For each POG, mean log₁₀-transformed relative abundance, sample prevalence, and log₁₀-transformed family size were calculated. Each variable was divided into ten quantile bins, and the combined abundance–prevalence–size labels defined matching strata. Within each stratum, non-CF-like POGs were randomly sampled without replacement to match the number of CF-like POGs in that stratum. This procedure was repeated 999 times using fixed random seeds, generating matched random sets with abundance, prevalence, and family-size distributions comparable to those of the observed CF-like POG set.

Empirical null models and randomization tests

Observed statistics from CF-like POG analyses were compared against empirical null distributions generated from 999 matched random POG sets. Empirical P values were calculated using one-sided tests in the predefined direction of the alternative hypothesis for each analysis. Standardized effect sizes (SES) were calculated using Eq. (1):

$$SES=\frac{S_{obs}-mean\left(S_{null}\right)}{s.d.\left(S_{null}\right)}$$ 1

where S_obs is the observed statistic and S_null denotes the null statistics obtained from matched random sets. Size-matched taxonomic randomization analyses used the same empirical null framework, with observed statistics calculated from CF-family models and null distributions generated from random species- or genus-level feature sets of the same size.

Assessment of taxonomic and phylogenetic structuring in CF-family profiles

Species-level CF-family presence/absence profiles were used to calculate Jaccard dissimilarities. Principal coordinates analysis (PCoA) was performed using the ape package (v5.8). Variation in CF-family profiles explained by taxonomic ranks was quantified using permutational multivariate analysis of variance (PERMANOVA) with 1000 permutations, implemented with the adonis2 function in the vegan package (v2.7-1). The R² values were adjusted for sample size and model degrees of freedom using the RsquareAdj function in the vegan package. Precomputed bacterial and archaeal species phylogenies from the UHGG catalog were used to calculate patristic distances with the ape package. Tree tips were matched to species with available CF-family presence/absence profiles, and the resulting patristic-distance matrices were paired with the corresponding Jaccard dissimilarity matrices. Because the bacterial species tree contained a large number of species and therefore generated a large number of pairwise distances, 0.01% of the nonredundant paired values from the bacterial patristic-distance and CF-profile dissimilarity matrices were randomly sampled. Spearman’s rank correlation was calculated for each subsample, and the procedure was repeated 20 times; the median correlation coefficient was reported. For archaea, all nonredundant paired distance values were used. For matched POG-based randomization analyses, the same analytical framework was applied to CF-like POG presence/absence profiles and each matched random POG set, allowing us to evaluate whether CF-like POGs exhibited stronger phylogenetic structuring than expected from matched protein-family backgrounds.

Definition of CF-based colonization strategies and genome-scale functional analysis

Species were clustered based on CF-family presence/absence profiles using Jaccard distances and partitioning around medoids (PAM), implemented with the pam function in the cluster package (v2.1.8.1). The optimal number of clusters was determined using the Calinski–Harabasz index, implemented in the clusterSim R package (v0.51-5). Cluster-specific CF families were identified by comparing the prevalence of each CF family across the resulting species-level clusters.

For each CF family, the numbers of species carrying that family in Cluster-1, Cluster-2, and Cluster-3 were denoted by c₁, c₂, and c₃, respectively. The corresponding total numbers of species in the three clusters were denoted by N₁, N₂, and N₃. Global differences in feature prevalence across clusters were first evaluated using Fisher’s exact test applied to the 3 × 2 contingency table shown in Eq. (2):

$$\left[\begin{array}{cc}\hfill c_1\hfill & \hfill N_1-c_1\hfill \\ \hfill c_2\hfill & \hfill N_2-c_2\hfill \\ \hfill c_3\hfill & \hfill N_3-c_3\hfill \end{array}\right]$$ 2

Cluster-specific enrichment was then assessed using three one-versus-rest Fisher’s exact tests. For example, enrichment of a given CF family in Cluster-1 was tested using the 2 × 2 contingency table shown in Eq. (3):

$$\left[\begin{array}{cc}\hfill c_1\hfill & \hfill N_1-c_1\hfill \\ \hfill c_2+c_3\hfill & \hfill \left(N_2+N_3\right)-\left(c_2+c_3\right)\hfill \end{array}\right]$$ 3

The same procedure was applied to Cluster-2 and Cluster-3. All Fisher’s exact tests were performed using Monte Carlo simulation with 10,000 replicates. P values were adjusted for multiple testing using the Benjamini–Hochberg (BH) method. The enrichment fold change for a given cluster was calculated as the CF-family prevalence in that cluster divided by its prevalence in the remaining two clusters combined. CF families with FDR < 0.05 and an enrichment fold change > 2 were considered significantly cluster-enriched.

To evaluate whether CF-defined clusters reflected broader genome-scale functional differences, KO annotations were extracted from eggNOG-mapper outputs and converted into species-level KO presence/absence profiles. Differentially prevalent KOs were identified using the same prevalence-testing framework described above. KEGG pathway enrichment analysis of cluster-enriched KOs was conducted using clusterProfiler (v4.17) and the official KEGG BRITE hierarchy files. Mantel tests were used to assess the association between CF-family and KO Jaccard dissimilarity matrices. Cluster-specific taxonomic enrichment at the phylum and family levels was also tested using the one-versus-rest Fisher’s exact test framework with BH correction.

Diversity analyses of CF abundance profiles

Alpha-diversity indices, including Shannon diversity and richness, were calculated from CF-gene abundance profiles. Richness was defined as the number of detected CF genes in each sample. For each case–control comparison, differences in alpha diversity were summarized as log₂-transformed fold changes between the case and control group means, and significance was assessed using Wilcoxon rank-sum tests followed by BH correction. For sample-level CF-gene abundance profiles, beta diversity was assessed using Bray–Curtis dissimilarity, and group differences were assessed using PERMANOVA with 1000 permutations. For matched randomization analyses, comparison-level alpha-diversity log₂ fold changes and PERMANOVA adjusted R² values were averaged across the 19 case–control comparisons and used as the observed statistics. Matched random POG profiles were analyzed using the same procedures and compared with observed CF-like POG statistics using empirical null models.

Concordance and PERMANOVA-based variance explanation analyses

Bray–Curtis dissimilarity matrices were calculated for CF-family abundance profiles and species-level taxonomic abundance profiles and used to generate PCoA ordinations. Procrustes analysis, followed by a permutation test with 1000 permutations, was performed to assess concordance between the two ordination configurations. Reciprocal analyses of variance-explanation were performed using PERMANOVA after dimensionality reduction. Specifically, principal component analysis (PCA) was applied separately to CF-family and species-level abundance matrices to reduce multicollinearity, and principal components cumulatively explaining 95% of the variance were retained as explanatory variables. This approach was used to estimate the fraction of species-level variation explained by CF-family profiles and, conversely, the fraction of CF-family profile variation explained by species-level composition. For CF-like POGs, the same Procrustes and PERMANOVA-based reciprocal analyses were applied and compared with matched random POG backgrounds using the empirical null model described above. To quantify the explanatory strength of individual CF families, the species-level Bray–Curtis dissimilarity matrix was used as the response, and the abundance profile of each CF family was fitted separately as an explanatory variable in PERMANOVA.

Meta-analysis of disease-associated CF families and microbial species

For each case–control comparison, feature-level effect sizes were calculated separately for CF-family and species-level abundance profiles. Standardized mean differences expressed as Hedges’ g and corresponding sampling variances were calculated using the escalc function from the metafor R package (v3.8-0). Random-effects meta-analyses were then performed using the rma function. Between-study heterogeneity was assessed using Cochran’s Q test and the I² statistic. P values from the random-effects models were adjusted for multiple testing using the BH method. Features with FDR < 0.05 were considered significantly associated with disease status.

CF-family carrier enrichment analysis

Species-level CF carrier profiles were derived from the species-level CF-family presence/absence matrix defined above. CF-family richness was calculated as the number of CF families present in each species. For richness and heterogeneity analyses, species with significant meta-analysis associations were classified as IBD-enriched or HC-enriched according to the sign of their species-level meta-analysis estimates. Wilcoxon rank-sum tests were used to compare total CF-family richness, as well as strategy-specific CF-family richness, between IBD-enriched and HC-enriched species. Within-group heterogeneity of CF carrier profiles was assessed using pairwise Jaccard dissimilarities calculated from total and strategy-specific CF-family presence/absence profiles.

CF-family carrier enrichment was assessed using a gene set enrichment analysis (GSEA)-like framework implemented with the GSEA function in the clusterProfiler package. All species were ranked by their meta-analysis-derived disease association estimates, with positive values indicating IBD bias and negative values indicating HC bias. For each CF family, the corresponding carrier species were treated as a species set and tested for enrichment toward either end of the ranked species list. CF families with positive normalized enrichment scores (NES) were interpreted as IBD-biased carrier sets, whereas those with negative NES were interpreted as HC-biased carrier sets. P values were adjusted across CF families using the BH method, and carrier sets with FDR < 0.05 were considered significant.

Leading-edge carrier species from significant CF carrier sets were extracted as core contributors to the enrichment signal. CF-family carrier relationships were visualized as a network in Cytoscape (v3.10.3), linking representative IBD- and HC-associated genera, their species, and CF families. Edges connected species to CF families when the corresponding CF family was present in that species. The direction of each CF-family association was determined by GSEA-like carrier enrichment, whereas the direction of each species association was determined by species-level meta-analysis estimates.

Random forest classification and feature-compression analyses

To evaluate the ability of CF-family signatures to discriminate IBD from HC status, we applied random forest classification, a commonly used approach for microbiome classification tasks⁵⁶. Models based on CF-family metagenomic abundance or metatranscriptomic expression profiles were trained using the randomForest package (v4.7) with 1000 trees. Within each dataset, 10-fold cross-validation was performed with fixed random seeds, and model performance was assessed by the area under the receiver operating characteristic curve (AUC), calculated from predicted class probabilities using the pROC package (v1.19). For cross-cohort prediction, models were trained on one dataset and evaluated on each of the remaining datasets.

The same random forest pipeline was applied to CF-like POG profiles and matched random POG sets for benchmarking. The average AUC across cohorts was used as the observed statistic and compared with the empirical null distribution generated from matched random POG sets. Inter-feature redundancy was quantified as the mean absolute Spearman correlation between pairs of features within each feature set.

To identify compact feature panels, we performed a post hoc feature-compression analysis. Feature importance was estimated using the mean decrease in accuracy (MDA) from cohort-specific random forest models. For each feature, MDA scores were averaged across cohorts, and CF families or CF-like POGs were ranked according to their mean cross-cohort importance. For each top-N feature set, random forest models were trained and evaluated using 10-fold cross-validation repeated five times with different random seeds within each cohort, and the cohort-specific AUCs were calculated as the mean across repeats.

To benchmark CF-family models against taxonomic feature sets of the same dimensionality, size-matched species- and genus-level benchmark analyses were performed. First, random species- and genus-level feature sets containing the same numbers of features as the corresponding CF-family models were generated 999 times, and the resulting AUC distributions were used as empirical null backgrounds. Second, species-level, genus-level, and CF-family features were ranked separately according to random forest variable importance, and the performance of classifiers based on the top-ranked 13, 46, and 79 features from each feature type was compared. One-sided empirical P values and SES values were calculated by comparing observed CF-family AUCs with the corresponding random taxonomic null distributions to test whether CF-family models outperformed size-matched random taxonomic feature sets.

Statistical analysis and visualization

All statistical analyses were performed using R (v4.4.3). Differential abundance analyses for each IBD cohort were conducted using the DESeq2 R package (v1.46) based on CF-family read counts. For non-parametric comparisons, the Kruskal–Wallis test was employed for multigroup analyses, followed by pairwise Wilcoxon rank-sum tests with BH correction. Cramér’s V was computed to assess the association between taxonomic distribution and CF-defined cluster assignments using the rstatix package (v0.7.2). Ternary plots were generated using the ggtern package (v3.5). General data visualization was performed using the ggplot2 (v3.5.2) and ggpubr (v0.6.1) R packages.

Reporting summary

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

Author contributions

S.J.Z. and J.-X.M. contributed to the conception and design of the study. J.-X.M., W.-D.L. and W.-F.T. downloaded and processed the publicly available datasets. J.-X.M. and W.-D.L. performed the bioinformatics analyses. J.-X.M., Y.Z. and S.J.Z. wrote the manuscript. W.-D.L., W.-F.T., Y.Z., J.G., S.T., X.Z., J.L. and J.W. contributed meaningful discussions. C.L. commented on and revised drafts of the manuscript. All authors were involved in preparing the manuscript and contributed to manuscript revision, reading, and approving the submitted version.

Peer review

Peer review information

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

Funding

This work is supported by grants from the National Natural Science Foundation of China (No. 82572567 and No. 323B2007) to S.J.Z. and Y.Z.

Data availability

The metagenomic and metatranscriptomic datasets used in this study are publicly available in the European Nucleotide Archive under the following accession codes PRJNA562600, PRJNA400072, PRJNA385949, PRJEB15371, PRJNA993675, PRJNA398089, PRJNA429990, PRJEB67456, PRJNA389280 and PRJNA436359. Detailed information for each dataset, including cohort name, data type, sample size, and accession code, is provided in Supplementary Data 12. The 27,096 reference colonization factor protein sequences used in this study were obtained from the genotype–habitat association project (GHA). The UHGG and UHGP sequence catalogs used in this study are publicly available from the Unified Human Gastrointestinal Genome (UHGG) catalog (v2.0.2) and the Unified Human Gastrointestinal Protein (UHGP) catalog (v2.0.2), respectively. The human reference genome GRCh38, used for host-read removal, is available under accession code GCF_000001405.40. All other data supporting the findings of this article are available in this article and its Supplementary Information. Source data are provided with this paper.

Code availability

The analysis code and supporting data associated with this study are publicly available on GitHub at https://github.com/mengjx855/25-CF-IBD-MGX and have been permanently archived on Zenodo under https://doi.org/10.5281/zenodo.21643082⁵⁷.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-76763-2.