Microbiome features associated with persistent intestinal carriages of Escherichia coli ST131 in a Southeast Asian cohort study Authors: Adrian Low, Zhuoya Yang, Kylin Treruangrachada Anantaya, Siyan Zhao, Wei Cong Tan, Rebecca Lynn Perez, Hao Chung The, Shirmin Zhi Yan Lim, Lin Liu, Jean-Sebastien Gounot, Junmei Samantha Kwah, Rick Twee-Hee Ong, Niranjan Nagarajan, Jonathan Wei Jie Lee, Yin Mo Published: 2026 ABSTRACT Escherichia coli sequence-type 131 (ST131) is the dominant global extraintestinal pathogen capable of asymptomatic intestinal carriage and sustained household transmission, challenging infection control. Despite its clinical significance, the ecological determinants of gut persistence remain poorly understood. We performed shotgun metagenomics on fecal samples to investigate gut microbiome features associated with ST131-positive samples, distinct host carrier statuses (persistent, intermittent and non-carriers) and household risks in a study of a Southeast Asian cohort. Here, we show that ST131 carriage was associated with compositional shifts without reducing species alpha-diversity. Regression analyses identified depletion of commensal taxa and the 1,5-anhydrofructose degradation pathway in ST131-positive samples. Persistent carriers exhibited highly perturbed microbiome enriched with pathobionts, aerobactin- and lipopolysaccharide (LPS)-biosynthesis pathways. Comparing household risk groups to control, revealed that biotin biosynthesis and 1,5-anhydrofructose degradation may influence ST131 co-colonization through both direct and indirect mechanisms. Machine learning analyses identified metabolic pathways as stronger discriminators of persistent carriage than taxonomic features. Genomic-resolved analysis of clinical ST131 isolates revealed conserved genes for iron-acquisition, LPS and antibiotic resistance determinants. Overall, while commensals and metabolism may influence initial ST131 colonization, persistent carriage is associated with specific microbial and metabolic adaptations, providing potential targets to limit intestinal ST131 persistence. Introduction The human gut microbiota is a complex and dynamic ecosystem shaped by host physiology, environmental exposures, and microbial interactions1. A central function of this community is colonization resistance, whereby resident microbes limit the expansion of opportunistic and pathogenic organism2. This protection is mediated through multiple, often overlapping mechanisms, including competition for nutrients3,4, inhibitory effects of microbial metabolites such as short-chain fatty acids5,6, production of antimicrobial compounds7,8, and modulation of host immune responses9. Despite these defenses, certain clinically important pathogens are still able to establish and persist within the intestinal niche, most notably Escherichia coli sequence type 131 (ST131). E. coli ST131 is the predominant extraintestinal pathogen (ExPEC) lineage, accounting for a high proportion of urinary tract infections (UTI) caused by uropathogenic E. coli (UPEC), and bacteremia worldwide10–12. ST131 is distinguished by its extensive multidrug resistance, and its capacity for prolonged, often asymptomatic gastrointestinal colonization, which provides a reservoir for person-to-person transmission within households and communities13–15. Although the antimicrobial resistance determinants and population structure of ST131 have been extensively characterized, including its association with carbapenem-resistant Enterobacteriaceae10,16,17, comparatively few studies have explored the gut microbial community composition and functional landscape that permits its intestinal colonization. One prior US-based study investigating ST131 within household co-residents identified taxonomic signatures associated with carriage of ST131-H30R subclone, including enrichment of phylum Actinomycetota, and genus Collinsella18. While informative, these observations were derived from 16S rRNA gene-based profiling, which lack the taxonomic resolution and functional depth required to resolve species-level variation or microbial metabolic traits relevant to colonization dynamics. In Southeast Asia, clinical surveillance has documented a high prevalence of ciprofloxacin-resistant ST131 C2/H30Rx subclones harboring the blaCTX-M-15 gene, yet the ecological context of these strains within the gut microbiome remains poorly defined12,19. A recent prospective study by Pérez et al examined ST131 transmission dynamics in this region and demonstrated that most carriers were intermittent carriers, with a minority exhibiting persistent carriage20. Building on this cohort, this study integrates shotgun metagenomic profiling of the gut microbiome with detailed carrier phenotyping, stratifying individuals by ST131 detection status (positive vs negative), host carriage pattern (persistent, intermittent and non-carriers) and household transmission risks (high, low and control). This study is in line with One Health Framework, through the integration of epidemiology and metagenomics to elucidate the gut microbiome configurations associated with ST13121. Through this approach, we identify taxonomic and functional features of the gut microbiome that implicate both host intestinal ecology and microbial mechanisms facilitating ST131 colonization and prolonged intestinal persistence. Results Gut microbiota of ST131-positive samples is compositionally different from ST131-negative samples To study gut microbiomes associated with ST131 detection, host carrier status, and household risk levels, reference-based metagenomic analysis was performed on 255 stool samples. Statistical associations were assessed using univariate and multivariate linear mixed-effects models, treating subjects as repeated measures to account for within-subject correlation and missing data (Fig. 1a). The mean age of participants was 42 ± 26.2 years, 60.5% or 72 females, 39.5% or 47 males, and 37.8% of the cohort had prior antibiotic exposure. Based on prior screening20, 14.1% (n = 36) of the stool samples had at least one ST131 isolate (ST131-positive), whereas 85.9% (n = 219) were ST131-negative. Samples were further stratified by host carrier status, comprising persistent carriers (9 individuals, n = 22 samples), intermittent carriers (30 individuals, n = 83 samples), non-carriers (80 individuals, n = 150 samples) (Fig. 1a).Fig. 1Study design and baseline gut microbiome diversity.a Study design flow chart illustrating metadata source, primary study groups and downstream analyses. Statistical test is based on two-sided Wilcoxon rank-sum test. b Boxplots showing the inverse Simpson’s index between ST131-positive and -negative samples and (c) among individuals with different carrier status. Each box represents the 25th and 75th percentiles, line inside box represents the median, and whiskers extend to the most extreme values within 1.5× interquartile range (IQR). d Principal coordinate analysis plot showing the beta-diversity between ST131-detection groups and (e) baseline carrier status. The percentage variance of each component (PC) axis is shown in parenthesis. Study design and baseline gut microbiome diversity. a Study design flow chart illustrating metadata source, primary study groups and downstream analyses. Statistical test is based on two-sided Wilcoxon rank-sum test. b Boxplots showing the inverse Simpson’s index between ST131-positive and -negative samples and (c) among individuals with different carrier status. Each box represents the 25th and 75th percentiles, line inside box represents the median, and whiskers extend to the most extreme values within 1.5× interquartile range (IQR). d Principal coordinate analysis plot showing the beta-diversity between ST131-detection groups and (e) baseline carrier status. The percentage variance of each component (PC) axis is shown in parenthesis. Alpha-diversity, assessed using the inverse Simpson’s index, did not differ between ST131-positive and ST131-negative samples (9.11 ± 6.04 vs. 9.85 ± 6.89, P-value = 0.66, Fig. 1b). Similarly, baseline alpha-diversity was comparable across persistent, intermittent and non-carriers (Fig. 1c). In contrast, beta-diversity based on the Bray-Curtis dissimilarity revealed a significant compositional separation between ST131-positive and ST131-negative samples (PERMANOVA, R2 = 0.009, P-value < 0.01) (Fig. 1d). This difference was further supported by PERMANOVA (R2 = 0.022, P-value < 0.05) using sample-size-matched data (n = 36 per group) (Supplementary Fig. 1a). While this single-iteration subsampling addresses group imbalance, it may not capture the full variance of the larger negative cohort. No significant differences in baseline beta diversity were observed when samples were stratified by host carrier status (PERMANOVA, R2 = 0.008, P-value = 0.455) (Fig. 1e). Multivariate analysis of baseline beta-diversity identified households (PERMANOVA, R2 = 0.316, P-value < 0.05) and prior antibiotic use (PERMANOVA, R2 = 0.0178, P-value < 0.05) as significant contributors to community variation, whereas age, sex and host carriage status were not associated with compositional differences (Suppl. Fig. 1b). Collectively, these findings indicate that the overall community structure is more strongly shaped by shared household environment and antibiotic exposure than by carriage phenotype alone. 1,5-anhydrofructose degradation pathway depleted in ST131-positive samples Differential abundance analysis of 255 fecal metagenomes, based on the species-level relative abundance generated from MetaPhlAn 4, was performed to identify taxa associated with ST131 status. Two Lachnospiraceae related species, Sellimonas intestinalis and an uncultured Lachnospiraceae bacterium_BX3, were significantly enriched in ST131-positive samples (q-value < 0.05). In contrast, five commensals Allisonella histaminiformans, Blautia wexlerae, Collinsella aerofaciens, Dysosmobacter sp., and Lachnospira pectinoschiza were significantly (q-value < 0.05) depleted in the ST131-positive samples (Fig. 2a). After adjusting for host carrier status, age, sex, and antibiotic exposure during the study period, no species was statistically significant.Fig. 2Differentially abundant microbial species and functional pathways between ST131-detection groups.a Barplot showing microbial species enriched or depleted in ST131-positive samples. b Metabolic pathways enriched or depleted in ST131-positive samples. All species and pathways shown are statistically significant (q-value < 0.05; log-fold change > 0.3) adjusted for subject-level random effects. Statistically significant taxa and pathways after adjusting for clinical and demography factors are shown in bold. Statistical results can be found in Zenodo archive92. c Boxplots showing log2 transformed copies per million (CPM) of 1,5-anhydrofructose degradation pathway in ST131-detection groups and antibiotics-naïve stratified samples. Each box represents the 25th and 75th percentiles, line inside box represents the median, and whiskers extend to the most extreme values within 1.5 × interquartile range (IQR). Asterisks indicate statistical significance at P < 0.05 based on two-sided Wilcoxon rank-sum test. Differentially abundant microbial species and functional pathways between ST131-detection groups. a Barplot showing microbial species enriched or depleted in ST131-positive samples. b Metabolic pathways enriched or depleted in ST131-positive samples. All species and pathways shown are statistically significant (q-value < 0.05; log-fold change > 0.3) adjusted for subject-level random effects. Statistically significant taxa and pathways after adjusting for clinical and demography factors are shown in bold. Statistical results can be found in Zenodo archive92. c Boxplots showing log2 transformed copies per million (CPM) of 1,5-anhydrofructose degradation pathway in ST131-detection groups and antibiotics-naïve stratified samples. Each box represents the 25th and 75th percentiles, line inside box represents the median, and whiskers extend to the most extreme values within 1.5 × interquartile range (IQR). Asterisks indicate statistical significance at P < 0.05 based on two-sided Wilcoxon rank-sum test. At the functional level, pathway-level analysis revealed ST131-positive samples were enriched in D-xylose degradation (PWY-7294) and formaldehyde assimilation II (RuMP cycle; PWY-1861), while 1,5-anhydrofructose degradation (PWY-6992) and anaerobic energy metabolism (PWY-7389) were depleted (q-value < 0.05) (Fig. 2b). Upon adjustment for clinical and demographic covariates, only 1,5-anhydrofructose degradation (PWY-6992) was differentially abundant with significantly lower abundance in the ST131-positive samples compared to the ST131-negative samples (Fig. 2c). A sensitivity analysis of antibiotic-naive samples validated this finding, suggesting that a reduced capacity for the 1,5-anhydrofructose degradation pathway is consistent with ST131 detection (Fig. 2c). 1,5-Anhydrofructose degradation and biotin biosynthesis pathways modulate ST131 co-colonization risk within households As individuals of the same households shared similar microbiomes based on Aitchison distance, we hypothesized that the risk of transmission would be greater within households with persistent carriers than in those without. Therefore, individuals were stratified into three-risk groups, high-risk (individuals in shared households with one or more persistent carriers), low-risk (individuals in shared households with one or more intermittent carriers), and control (only non-carriers). Differential abundance analysis identified distinct microbial signatures across the risk groups (Fig. 3a). While most enriched species differed between the high- and low-risk groups, four species (Roseburia sp. AF02_12, Clostridales bacterium KLE1615, Bifidobacterium catenulatum, Anaerosacchariphilus sp. NSJ 68) were consistently enriched in both groups relative to controls. Prevotella sp. 885 was uniquely enriched in high-risk households compared to low-risk and control groups (Fig. 3a). Conversely, a consistent depletion of Enterocloster clostridioformis and Erysipelotrichiaceae bacterium 3_1_53 was observed in both risk groups compared to controls (Fig. 3a). However, the prevalence of Erysipelotrichiaceae bacterium 3_1_53 was relatively lower than E. clostridioformis in the controls (Suppl. Fig. 2). Functional profiling identified just two key metabolic pathways that distinguished the household risk groups from controls (Fig. 3b). Specifically, 1,5-anhydrousfructose degradation (PWY-6992) was significantly depleted in high-risk households compared to controls (Fig. 3b). Conversely, biotin biosynthesis II (PWY-5005) was significantly enriched in low-risk group relative to controls (Fig. 3b).Fig. 3Differential abundance analysis of taxonomic and metabolic pathways across the household risk groups.a Heatmap showing the mean relative abundance of differentially abundant species across high-risk, low-risk and control groups. Significant taxa were identified using univariate linear regression comparing high- or low-risk groups relative to control (estimate > 0.3, q-value < 0.05) as indicated by asterisks. Statistical results can be found in Zenodo archive92. Unsupervised hierarchical clustering based on Euclidean distance was applied to taxa and risk groups. Species names are color coded by the group in which they are enriched. Black text indicates taxon with significant shifts across high- and low-risks groups. Boxplot distributions for these data-points are shown in Supplementary Fig. 2. b Log2 transformed copies per million (CPM) of two metabolic pathways. Global statistical test was determined using the Kruskal-Wallis test followed by two-sided Dunn’s post-hoc test for pairwise comparison. Asterisks denote statistical significance, ‘*’ denotes q-value < 0.05, ‘**’ denotes q-value < 0.01 and ‘***’ denotes q-value < 0.001. Each q-value represents P-value adjusted using Benjamini-Hochberg false-discovery rate (FDR) correction. Dunn’s test q-values for PWY-6992 (Control vs High Risk: 0.0039, Control vs Low Risk: 0.5242, High Risk vs Low Risk: 0.0147), q-values for PWY-5005: (Control vs High Risk: 0.0333, Control vs Low Risk: 0.0001, High Risk vs Low Risk: 0.0333). Each box represents the 25th and 75th percentiles, line inside box represents the median, and whiskers extend to the most extreme values within 1.5× interquartile range (IQR). Differential abundance analysis of taxonomic and metabolic pathways across the household risk groups. a Heatmap showing the mean relative abundance of differentially abundant species across high-risk, low-risk and control groups. Significant taxa were identified using univariate linear regression comparing high- or low-risk groups relative to control (estimate > 0.3, q-value < 0.05) as indicated by asterisks. Statistical results can be found in Zenodo archive92. Unsupervised hierarchical clustering based on Euclidean distance was applied to taxa and risk groups. Species names are color coded by the group in which they are enriched. Black text indicates taxon with significant shifts across high- and low-risks groups. Boxplot distributions for these data-points are shown in Supplementary Fig. 2. b Log2 transformed copies per million (CPM) of two metabolic pathways. Global statistical test was determined using the Kruskal-Wallis test followed by two-sided Dunn’s post-hoc test for pairwise comparison. Asterisks denote statistical significance, ‘*’ denotes q-value < 0.05, ‘**’ denotes q-value < 0.01 and ‘***’ denotes q-value < 0.001. Each q-value represents P-value adjusted using Benjamini-Hochberg false-discovery rate (FDR) correction. Dunn’s test q-values for PWY-6992 (Control vs High Risk: 0.0039, Control vs Low Risk: 0.5242, High Risk vs Low Risk: 0.0147), q-values for PWY-5005: (Control vs High Risk: 0.0333, Control vs Low Risk: 0.0001, High Risk vs Low Risk: 0.0333). Each box represents the 25th and 75th percentiles, line inside box represents the median, and whiskers extend to the most extreme values within 1.5× interquartile range (IQR). Enrichment of aerobactin and lipopolysaccharide biosynthesis pathways in persistent carriers Analysis on host carrier status was performed to identify microbial markers that may increase the susceptibility of ST131 intestinal carriage. Five species were significantly (q-value < 0.05) enriched in persistent carriers relative to non-carriers: Enorma massiliensis, a close relative of Collinsella in the family Coriobacteriaceae, Kluyvera genomosp_3, a pathobiont in the same Enterobacteriaceae family as E. coli, Clostridium perfringens, a pathobiont, Intestinimonas timonensis, a commensal of the family Oscillospiraceae, and an uncharacterized Scatosoma sp. (Fig. 4a). The only depleted species was Romboutsia timonensis (q-value < 0.05). After adjustment for ST131-detection, age, gender, and antibiotic exposure, C. perfringens, I. timonensis, and R. timonensis remained significant, underscoring their robust association with persistent carriers.Fig. 4Differential abundance analysis of microbial species and functional pathways between carrier status-based groups.a A cladogram showing the statistically significant (q-value < 0.05) differentially abundant bacterial species for each pair-wise comparison. Family level taxonomic assignment is shown in parenthesis. b Prevalence and relative abundance of key differentially abundant species in persistent carriers using multivariate analysis. c Boxplots highlighting the significant differentially abundant pathways between carrier status-based groups (non-carrier, intermittent, and persistent). MetaCyc identities are shown preceding pathway names. Each box represents the 25th and 75th percentiles, line inside box represents the median, and whiskers extend to the most extreme values within 1.5× interquartile range (IQR). Statistically significant pathways (q-value < 0.05) after multi-variate analysis are shown in bold. d Prevalence and relative abundance of key differentially abundant pathways across carrier status-based groups identified after multivariate analysis. Asterisks denote statistical significance where ‘*’ denotes q-value < 0.05, ‘**’ denotes q-value < 0.01 and ‘***’ denotes q-value < 0.001 based on two-sided Dunn’s post-hoc test following Kruskal-Wallis test. Linear regression results and statistical results with exact P-values can be found in Zenodo archive92. Differential abundance analysis of microbial species and functional pathways between carrier status-based groups. a A cladogram showing the statistically significant (q-value < 0.05) differentially abundant bacterial species for each pair-wise comparison. Family level taxonomic assignment is shown in parenthesis. b Prevalence and relative abundance of key differentially abundant species in persistent carriers using multivariate analysis. c Boxplots highlighting the significant differentially abundant pathways between carrier status-based groups (non-carrier, intermittent, and persistent). MetaCyc identities are shown preceding pathway names. Each box represents the 25th and 75th percentiles, line inside box represents the median, and whiskers extend to the most extreme values within 1.5× interquartile range (IQR). Statistically significant pathways (q-value < 0.05) after multi-variate analysis are shown in bold. d Prevalence and relative abundance of key differentially abundant pathways across carrier status-based groups identified after multivariate analysis. Asterisks denote statistical significance where ‘*’ denotes q-value < 0.05, ‘**’ denotes q-value < 0.01 and ‘***’ denotes q-value < 0.001 based on two-sided Dunn’s post-hoc test following Kruskal-Wallis test. Linear regression results and statistical results with exact P-values can be found in Zenodo archive92. Comparison between persistent and intermittent carriers revealed three enriched species in persistent carriers, Clostridium perfringens, Eubacterium sp. AF16_48, and Kluyvera genomosp_3 (q-value < 0.05), and two depleted species, Eubacterium ventriosum and R. timonensis (q-value < 0.05). After adjusting for clinical covariates, Eubacterium sp. AF16_48 and R. timonensis remained differentially abundant. In the intermittent versus non-carrier comparison, five taxa were enriched (q-value < 0.05) including Anaerostipes caccae, Lacrimispora saccharolytica and Desulfovibrio_SGB5076 (Fig. 4a), while Bifidobacterium bifidum and Collinsella aerofaciens were found to be depleted (q-value < 0·05); none remained significant after adjustment. The prevalence and relative abundance patterns of the identified species were consistent with the differential findings; for example, I. timonensis was more prevalent and with significantly greater relative abundance in persistent carriers compared to intermittent and non-carriers (Fig. 4b), whereas R. timonensis was less prevalent and lower in relative abundance in persistent carriers compared to the other carrier status (Fig. 4b). At the functional level, eight metabolic pathways were enriched in persistent carriers versus non-carriers, including aerobactin biosynthesis (AEROBACTINSYN-PWY), lipopolysaccharide biosynthesis (LPSSYN-PWY), two ubiquinol-8 biosynthesis pathways (PWY-5180 and PWY-5415), a pathway for phospholipid remodeling (PWY-7409), and D-xylose degradation IV (PWY-7294). No pathways were depleted. Following adjustment for ST131 detection, age, gender and prior antibiotic use, aerobactin (AEROBACTINSYN-PWY), and lipopolysaccharide (LPSSYN-PWY) biosynthesis pathways remained differentially abundant (Fig. 4d). Comparison between persistent and intermittent carriers identified 22 significantly differentially abundant pathways (linear regression, q-value < 0.05), 21 enriched and one depleted, where seven of the enriched pathways overlapped with those identified with persistent versus non-carrier comparison. Methanogenesis from acetate (METH-ACETATE-PWY) was the single depleted pathway. Aerobactin and lipopolysaccharide biosynthesis pathways remained robustly enriched in persistent carriers after adjusting for clinical and demographic variables, and were present in ~50% of persistent carriers with higher relative abundance compared to intermittent and non-carriers (Fig. 4d). Phosphatidate metabolism (PWY-7039) was the only pathway (q-value < 0.05) depleted in intermittent carriers compared to non-carriers but was not significant after adjustment. Collectively, persistent carriers exhibit taxonomic signatures (e.g., Clostridium perfringens, Kluyvera and Intestinimonas timonensis) associated with dysbiosis and metabolic mechanisms (e.g., aerobactin, lipopolysaccharide) that may facilitate ST131 persistence. A sensitivity analysis of antibiotics naïve samples specifically of C. perfringens, I. timonensis, and R. timonensis and Eubacterium sp. AF16_48 and two metabolic features for aerobactin, lipopolysaccharide biosynthesis largely support the directional enrichment of the features, except for I. timonensis where the prevalence within the persistent group is relatively low ( ~ 13%) (Suppl. Fig. 3). Aerobactin genes strongly correlated with Escherichia coli relative abundance in persistent carriers Comparison of E. coli relative abundance between ST131-positive and ST131-negative samples revealed no significant difference, whereas significant differences were observed across host carrier status groups (Suppl. Fig. 4, Suppl. Fig. 5). To identify potential function biomarkers associated with E. coli abundance across carrier states, we examined the aerobactin biosynthesis pathway at the genes level. This pathway comprises four enzymes encoded by iucD (L-lysine N6-monooxygenase (NADPH), EC 1.14.13.59), iucB (N6-hydroxylysine N-acetyltransferase, EC 2.3.1.102), iucA (N2-citryl-N6-acetyl-N6-hydroxylysine synthase, EC 6.3.2.38), and iucC (aerobactin synthase, EC 6.3.2.39) (Fig. 5a)22–24. Spearman correlation analysis demonstrated strong correlations between the abundance of all four aerobactin genes and E. coli in persistent carriers (rhopersistent = 0.811–0.852), compared with weak to moderate correlations in intermittent (rhointermittent = 0.276–0.371) and non-carriers (rhonon-carriers = 0.395–0.485) (Fig. 5b). Species-stratified analysis revealed E. coli as the most prevalent and dominant species harboring the four genes, followed by Klebsiella pneumoniae and other members of the Enterobacteriaceae (Supplementary Fig. 6). While previous study has identified iucD as potential marker for ST131 carriage10, our results show that E.coli specific iucA stood out to be significantly more abundant than K. pneumoniae especially in persistent carriers (Suppl. Fig. 7). However, this was not the case with iucBCD as K. pneumoniae had similar relative abundances of those genes compared to E.coli (Suppl. Fig. 7).Fig. 5Associating Escherichia coli relative abundance with genes involved in the aerobactin biosynthesis pathway among carrier status.a Enzymes involved in aerobactin biosynthesis pathway. b Spearman correlation of the four genes to Escherichia coli relative abundance. Solid lines indicate linear lines of best fit for each group, and the surrounding shaded error bands represent the 95% confidence intervals. Asterisks denote statistical significance where ‘*’ denotes P-value < 0.05, ‘**’ denotes P-value < 0.01 and ‘***’ denotes P-value < 0.001. Non-carriers (n = 150), Persistent (n = 22), Intermittent (n = 83). Statistical results with exact P-values can be found in Zenodo archive92. Associating Escherichia coli relative abundance with genes involved in the aerobactin biosynthesis pathway among carrier status. a Enzymes involved in aerobactin biosynthesis pathway. b Spearman correlation of the four genes to Escherichia coli relative abundance. Solid lines indicate linear lines of best fit for each group, and the surrounding shaded error bands represent the 95% confidence intervals. Asterisks denote statistical significance where ‘*’ denotes P-value < 0.05, ‘**’ denotes P-value < 0.01 and ‘***’ denotes P-value < 0.001. Non-carriers (n = 150), Persistent (n = 22), Intermittent (n = 83). Statistical results with exact P-values can be found in Zenodo archive92. Enrichment of sulfonamide, macrolide, aminoglycosides and beta-lactam resistance genes across ST131 carriage and persistent carriers ST131-positive samples exhibited significantly higher antibiotic resistance gene (ARG) richness compared to ST131-negative samples (Fig. 6a). In contrast, no significant difference in ARG richness was observed across host carrier status groups at baseline (Fig. 6b). Univariate differential abundance analysis identified five ARGs (aadA5, Mrx, mphA, sul1, and OmpK35) that were significantly enriched in ST131-positive samples (q-value < 0.05), conferring resistance to aminoglycosides, macrolides, sulfonamides, and beta-lactams (Fig. 6c). After adjustment for host carrier status, age, gender and prior antibiotic exposure, four of these ARGs remained significantly enriched, except for mphA.Fig. 6Diversity and differential abundance analysis of antibiotic resistance genes (ARGs).a ARG richness across ST131 detection and b baseline host carrier status. Each box represents the 25th and 75th percentiles, line inside box represents the median, and whiskers extend to the most extreme values within 1.5× interquartile range (IQR). c Differential abundance analysis between ST131 detection groups, and between carrier status groups. CARD accession numbers are shown in parentheses. Genes with resistance to multiple-drug classes are shown as ‘+’. ARG fonts are colored to match resistant phenotype. Asterisk denotes genes that are statistically significant (q-values < 0.05) after adjusting for clinical and demographic factors. Linear regression statistical results with exact P-values can be found in Zenodo archive92. Diversity and differential abundance analysis of antibiotic resistance genes (ARGs). a ARG richness across ST131 detection and b baseline host carrier status. Each box represents the 25th and 75th percentiles, line inside box represents the median, and whiskers extend to the most extreme values within 1.5× interquartile range (IQR). c Differential abundance analysis between ST131 detection groups, and between carrier status groups. CARD accession numbers are shown in parentheses. Genes with resistance to multiple-drug classes are shown as ‘+’. ARG fonts are colored to match resistant phenotype. Asterisk denotes genes that are statistically significant (q-values < 0.05) after adjusting for clinical and demographic factors. Linear regression statistical results with exact P-values can be found in Zenodo archive92. Across carrier phenotypes, sul1, emrE and gadW were consistently enriched in persistent carriers with intermittent carriers and non-carriers (Fig. 6c). Notably, gadW, encoding a multidrug efflux regulator associated with resistance to beta-lactam and macrolide, remained significantly enriched in persistent carriers after adjustment for clinical and demographic factors (Fig. 6c). In contrast, multiple tetracycline genes, including tet(32) encoding a ribosomal protection protein, were significantly depleted in persistent carriers relative to intermittent carriers and non-carriers (Fig. 6c). The trimethoprim resistance gene dfrA17 was enriched in persistent carriers compared to non-carriers but not when compared with intermittent carriers (Fig. 6c). Fluoroquinolone resistance genes QnrS5 and florR were also enriched in persistent carriers; however, these associations were no longer significant after covariate adjustment. Collectively, these analyses indicate that ST131 carriage, particularly persistent carriage, is associated with an expanded and distinct resistome, characterized by enrichment of multidrug resistance determinants, whereas ARG richness alone does not differentiate carrier phenotypes. Conserved multiple siderophore acquisition types, and macrolide resistance genes among clinically derived STs Whole-genome assemblies (WGA) were generated for six E. coli strains isolated from urine or blood samples of index patients with urinary tract infections or bacteremia, achieving high completeness (mean 99.97%) with low contamination ( ≤ 1.53%) (Table 1, Supplementary Table 1). In silico multi-locus typing (MLST) identified five ST131 strains and one ST73 strain, corresponding to serotypes O16:H5 or O25:H4 (Supplementary Table 1). These genomes were analyzed to characterize metabolic potential and antibiotic resistance determinants of clinically invasive E. coli sequence types. All strains encoded multiple siderophore reuptake systems, including a conserved enterobactin biosynthesis and transport pathway (Table 1). Aerobactin biosynthesis genes (iucABCD), and the aerobactin receptor gene (iutA) were conserved in three of the five ST131 isolates and in the ST73 isolate (Table 1). Plasmid replicons from the IncL, IncF, and Col groups were detected in all ST131 isolates, whereas the ST73 isolate harbored only a single Col (RNAi) plasmid ori.Table 1Siderophore systems in five ST131 and one ST73 strainStrain ID (Biosample)STSiderophore Biosynthesis GenesSiderophore Receptor GenesAerobactin genes iucC/D/A/BEnterobactin synthase genes (entB/E/F)Aerobactin receptor, (iutA)Siderophore transport system permease protein yfiZ/yusV iron-hydroxamate siderophores schizokinenCatecholate siderophore receptor gene fiuEnterobactin transport system permease protein genes fepD/G/C/EProtein TonB /Biopolymer transport protein genes exbB/exbDEC51 (urine)131✗✓✗✓✓✓✓EC57 (urine)131✓✓✓✓✓✓✓EC74 (blood)73✓✓✓✓✓✓✓EC77 (urine)131✗✓✗✓✓✓✓EC81 (urine)131✓✓✓✓✓✓✓EC85 (blood)131✓✓✓✓✓✓✓✓ present or ✗ absent in the draft genome. Table 1Siderophore systems in five ST131 and one ST73 strainStrain ID (Biosample)STSiderophore Biosynthesis GenesSiderophore Receptor GenesAerobactin genes iucC/D/A/BEnterobactin synthase genes (entB/E/F)Aerobactin receptor, (iutA)Siderophore transport system permease protein yfiZ/yusV iron-hydroxamate siderophores schizokinenCatecholate siderophore receptor gene fiuEnterobactin transport system permease protein genes fepD/G/C/EProtein TonB /Biopolymer transport protein genes exbB/exbDEC51 (urine)131✗✓✗✓✓✓✓EC57 (urine)131✓✓✓✓✓✓✓EC74 (blood)73✓✓✓✓✓✓✓EC77 (urine)131✗✓✗✓✓✓✓EC81 (urine)131✓✓✓✓✓✓✓EC85 (blood)131✓✓✓✓✓✓✓✓ present or ✗ absent in the draft genome. Siderophore systems in five ST131 and one ST73 strain ✓ present or ✗ absent in the draft genome. Since the model 1,5-anhydrofrutose reductases (Afr, EC 1.1.1.292 and EC 1.1.1.263) are derived from Ensifer adhaerens and mammalian hosts, respectively, a direct ortholog has not been identified in E. coli25. Using a hidden Markov Model (HMM) trained KO database, we searched homologs within the genomes of the five ST131 strains, and two E. coli K12 substrains (C600 and MG1655) linked to the metabolism of 1,5-anhydrofructose or its downstream metabolites26,27. A group of highly conserved proteins (351 aa) were identified across all strains with close homology to the Afr (EC 1.1.1.292), although scores were below the HMM bit-score threshold (Suppl. Table 2). The putative Afr in the five ST131 strains shared high similarity (mean 98.4 ± 0.33%) with proteins from the Gfo/Idh/MocA oxidoreductase family found in substrains C600 and MG1655 (Suppl. Table 2). Furthermore, phylogenetic analysis confirmed these bacterial Afr candidates are more closely related to the E. adhaerens reference sequence than to the human Afr (EC 1.1.1.263) (Suppl. Fig. 8). A large proportion of ARGs were shared across ST131 isolates, with 76.8% (53/69) conserved among all ST131 genomes and only two ARGs differing ST131 and ST73. About half (50.9%) of these shared ARGs confer resistance to fluoroquinolones (30.2%), and macrolides (20.8%), including multiple multidrug resistance mechanisms. These resistance determinants were predominantly efflux pumps of the resistance-nodulation-cell division (RND) superfamily and the major facilitator superfamily, which mediate resistance to fluoroquinolones and other anti-microbial classes. Notably, two ARGs identified as differentially enriched in ST131-positive samples and persistent carriers - emrE (efflux pump) and gadW (transcriptional regulator) - were conserved across all sequenced isolates. Other ARGs associated with ST131 carriage, including gyrA, Mrx and sul1, were present in at least one isolate (Suppl. Fig. 9). The ST73 contained a beta-lactamase (EC-5) conferring resistance to cephalosporin, which was absent from all ST131 genomes. The extended-spectrum β-lactamase (ESBL) gene blaCTX-M-15, was detected in only two ST131 isolates (EC77 and EC81) (Suppl. Fig. 9). ARGs conferring tetracycline resistance were limited to efflux-based mechanisms (emrY, emrK, tetA) with no evidence of ribosomal protection proteins such as tet(32). Similarly, the trimethoprim gene dfrA17, was not identified in any of the sequenced isolates. Collectively, these genome-resolved analyses delineate a conserved resistome and iron acquisition repertoire characteristic of clinical ST131 isolates, supported by extensive ARGs with predominance of efflux-mediating resistance, and conserved siderophore systems that distinguish ST131 from the ST73 comparator strain. Microbial signatures provide additive predictive power for ST131 carriage beyond clinical factors Machine learning was applied to identify microbial species and metabolic pathways that best distinguish persistent carriers from intermittent carriers and non-carriers. Models incorporating either taxonomic or functional features achieved higher discriminatory performance than models based solely on clinical variables, as assessed by receiver operating characteristic (ROC) analysis in both comparisons (Figs. 7a, b). At the taxonomic level, machine learning identified a suite of 15 microbial taxa as the most robust predictors of persistent carriers from non-carriers (Fig. 7c). A notable divergence was observed between closely related species, where Collinsella intestinale was significantly enriched in persistent carriers, whereas Collinsella aerofaciens was significantly depleted (Fig. 7c, Suppl. Fig. 5). These predictive features diverged from taxa identified via linear regression (Fig. 7c). This suggests that reliable classification distinguishable between persistent and non-carriers may rely on taxa contributing through non-linear interactions rather than independent linear shifts in abundance. In contrast, Eubacterium sp. AF16_48 and Romboutsia timosensis were among the most informative taxa differentiating persistent carriers from intermittent carriers (Fig. 7d). At the functional level, pathways related to aerobactin biosynthesis and LPS biosynthesis ranked among the top features distinguishing persistent carriers from non-carriers and intermittent carriers (Fig. 7e, f). Importantly, these functional pathways and consensus taxonomic features were identified independently by both linear mixed-effect modelling and random forest analyses. This cross-method convergence provides internal validation of these features as candidate biomarkers of persistent ST131 co-colonization. Given the limited number of persistent carriers in the our dataset, these findings should be interpreted as exploratory and warrant validation in larger, prospective cohorts.Fig. 7Random forest regression models to identify taxa and pathway discriminators to distinguish persistent carriers from non-carriers and intermittent carriers.a Receiver operating characteristic curve (ROC) for persistent carriers vs non-carriers based on taxa, pathways and clinical factors. b Receiver operating characteristic curve (ROC) for persistent carriers vs intermittent carriers based on taxa, pathways and clinical factors. c Top 15 taxa distinguishing persistent carriers from non-carriers (ranked by Mean Decrease Gini) with corresponding boxplots showing relative abundances. d Top 15 taxa that distinguish between persistent carriers and intermittent carriers. e Top 15 pathways distinguishing persistent carriers from non-carriers (ranked by Mean Decrease Gini) with corresponding boxplots showing log2 copies per million (CPM) pathways. f Top 15 pathways distinguishing persistent carriers from intermittent carriers (ranked by Mean Decrease Gini) with corresponding boxplots showing log2 copies per million (CPM) pathways. Each box represents the 25th and 75th percentiles, line inside box represents the median, and whiskers extend to the most extreme values within 1.5× interquartile range (IQR). Taxa and pathways in red text indicate microbial features consistent with significant features based on differential abundance analysis adjusted for clinical factors. Asterisks denote statistical significance using two-sided Wilcoxon rank-sum test with Benjamini-Hochberg adjusted P-values (‘*’ q-value < 0.05, ‘**’ q-value < 0.01 and ‘***’ q-value < 0.001). Statistical results including exact P-values can be found in Zenodo archive92. Random forest regression models to identify taxa and pathway discriminators to distinguish persistent carriers from non-carriers and intermittent carriers. a Receiver operating characteristic curve (ROC) for persistent carriers vs non-carriers based on taxa, pathways and clinical factors. b Receiver operating characteristic curve (ROC) for persistent carriers vs intermittent carriers based on taxa, pathways and clinical factors. c Top 15 taxa distinguishing persistent carriers from non-carriers (ranked by Mean Decrease Gini) with corresponding boxplots showing relative abundances. d Top 15 taxa that distinguish between persistent carriers and intermittent carriers. e Top 15 pathways distinguishing persistent carriers from non-carriers (ranked by Mean Decrease Gini) with corresponding boxplots showing log2 copies per million (CPM) pathways. f Top 15 pathways distinguishing persistent carriers from intermittent carriers (ranked by Mean Decrease Gini) with corresponding boxplots showing log2 copies per million (CPM) pathways. Each box represents the 25th and 75th percentiles, line inside box represents the median, and whiskers extend to the most extreme values within 1.5× interquartile range (IQR). Taxa and pathways in red text indicate microbial features consistent with significant features based on differential abundance analysis adjusted for clinical factors. Asterisks denote statistical significance using two-sided Wilcoxon rank-sum test with Benjamini-Hochberg adjusted P-values (‘*’ q-value < 0.05, ‘**’ q-value < 0.01 and ‘***’ q-value < 0.001). Statistical results including exact P-values can be found in Zenodo archive92. Discussion In this prospective Southeast Asian cohort, the gut microbiomes of individuals previously screened for community-level ST131 transmission were characterized. The findings reveal microbiome features associated with ST131 intestinal carriage, carrier persistence, household exposure risks and host susceptibility, providing insight into the ecological and functional context that may facilitate ST131 colonization. Comparative analyses between ST131-positive and ST131-negative samples suggest that fecal carriage of ST131 is not accompanied by a global loss of microbial diversity, despite clear compositional differences. This contrasts with findings from a US-based cohort, where higher phylum-level alpha diversity was reported in ST131-positive samples18. In the same study, Collinsella was identified as a robust genus of persistent ST131 carriers18. This enrichment may be attributed to Collinsella intestinale and less C. aerofaciens as seen in the Southeast Asian cohort. Nevertheless, these microbial markers require further validation across diverse populations, as factors such as diet, and host genetics likely contribute to variations in gut microbiome composition28,29. Furthermore, discrepancies between cohorts may stem from differences in study design, where the US study captured long-term carriage dynamics over a multi-year period, whereas the current study applied a cross-sectional approach to the analysis18. Among species enriched in ST131-positive samples, Sellimonas intestinalis, previously shown to be negatively associated with ESBL-producing E. coli in intensive care unit patients30, was found to be associated with asymptomatic carriage of ST131. Conversely, depleted species including Blautia wexlerae and certain Dysosmobacter spp. are known producers of acetate and butyrate31,32, which can inhibit E. coli growth5. As microbial interaction is complex, recent studies suggest a combination low interference by inhibitory compounds such as antimicrobials or short-chain fatty acids coupled with weak resource competition (i.e., substrate metabolism specificity) may facilitate extraintestinal pathogenic strains like ST131 to successfully colonize the gut33. The observed depletion of the 1,5-anhydrofructose degradation pathway may lead to an accumulation of this monosaccharide, which has moderate effects on the gut microbial composition as seen in mouse model34, potentially creating an environment for E. coli ST131 to co-colonization. 1,5-anhydro-D-fructose catabolism and its by-products have been demonstrated in substrains of the model E. coli K1226,35. This pathway, including the downstream anhydroglycolysis pathway, represents a specialized metabolic route to utilize glycogen under glucose-limiting conditions26,27,35–37. Putative identification of this protein in this study suggests that this metabolic capability may be conserved across E. coli strains. Further characterization of gene regulation and kinetics of this putative 1,5-anhydrofrutose reductase is essential to confirm a functional role. Additionally, co-culturing experiments with gut commensals are warranted to determine if E. coli actively leverages this substrate as a mechanism for species displacement and niche dominance. Classification by household risk identified features that are independent of individual carrier status (e.g., individuals within a high-risk household may be persistent, intermittent, or non-carriers). While it is not clear how the two largely different species associated with high- and low-risk groups contribute to ST131 colonization susceptibility, Enterocloster clostridioformis associated with the control group has been shown to elicit host immunity to protect against Salmonella in murine gut38. The identification of only two metabolic pathways indicates functional redundancy within these communities. The depletion of 1,5-anhydrousfructose degradation in the high-risk group is consistent with ST131-positive samples, suggesting a gut environment that is permissive to ST131 colonization. The enrichment of biotin biosynthesis II pathway in low-risk group may influence E. coli ST131 co-colonization. Biotin levels modulate the expression of adherence-associated proteins required for intestinal colonization in related pathotypes, such as enterohemorrhagic E. coli, in a concentration-dependent manner39. Stratification by carrier status further revealed distinct microbiome features associated with persistent ST131 carriage. Persistent carriers were enriched with Clostridium perfringens, a species often associated with intestinal inflammation40. In contrast, intermittent carriers were enriched with butyrate-producing taxa such as Anaerostipes caccae, and Lacrimispora saccharolytica41, whereas sulfate-reducing bacteria such as Desulfovibrio sp., commonly linked to intestinal inflammation, were also detected, potentially reflecting a transient inflammatory state42. Functional profiling identified aerobactin and LPS biosynthesis pathways as robust biomarkers for persistence across multiple analytical frameworks. Iron is a tightly restricted resource in the mammalian gut due to host-mediated nutritional immunity, creating intense competition between microbes processing diverse siderophore systems43,44. Our results suggest that ST131 utilizes aerobactin-mediated iron acquisition to outcompete other microbes. Interestingly, among the four aerobactin genes, species-stratified analysis suggests that iucA may mark increased susceptibility to ST131 colonization, whereas iucB, iucC, and iucD may reflect co-colonization with other Enterobacteriaceae species. Similarly, structural variations in LPS particularly with the lipid A moiety can attenuate TLR4 signaling and suppress NF-κB activation, enabling immune evasion45–48. Given their central biological roles, both pathways represent attractive therapeutic targets against ExPEC, including ST13149,50. Genome-resolved analyses of clinical isolates confirmed that ST131 and the comparator ST73 both encode multiple siderophore systems including catechol (enterobactin), hydroxamate (aerobactin), and dihydroxamate (schizokinen)24,51 transporters. While ST131 metagenomes were enriched for ARGs like mphA, sul1, and aadA5, the clinical isolates showed a more varied profile. Notably, the ESBL gene blaCTX-M-15, prevalent in ST131 C2/H30Rx subclones causing bacteremia in Southeast Asia12,19, was rare in our clinical isolates. This observation aligns with prior findings from paired urine-fecal isolates lacking blaCTX-M-15 but harboring AmpC beta-lactamases17, suggesting that resistance profile may differ between UTI-associated and invasive ST131 strains. Broader sampling will be necessary to validate the distinction. Several limitations should be acknowledged. The relatively small number of persistent carriers constrains statical power and model generalizability. Consequently, the microbial features identified here warrant further validation in larger, independent cohorts. Another limitation is that short-read metagenomics has inherent limitations in resolving E. coli sequence types due to low abundance and limited discriminatory loci. Targeted approaches, such as custom probe capture or transcriptomic profiling, may improve strain-level resolution but introduce additional biases and are best suited for focused functional studies52. Despite these limitations, our study provides an integrated taxonomic, functional, and resistome-level characterization of ST131 intestinal carriage in an underrepresented Southeast Asian population, advancing understanding of the microbial and ecological determinants of ST131 persistence. In alignment with the One Health framework, integrating epidemiology and metagenomics is critical for mitigating the transmission of ExPEC. Through microbiome profiling, we identified iron acquisition and LPS biosynthesis as conserved metabolic pillars underpinning increased susceptibility to prolonged ST131 co-colonization. Our findings underscore the dominant influence of host, household exposure risk and environmental factors, particularly antibiotic exposure and household sharing, in shaping ST131-associated microbiome configurations, while revealing ecological and functional adaptations that may confer a competitive advantage to this globally disseminated lineage. Together, these results provide a foundation for developing microbiome-informed strategies to mitigate the transmission of ExPEC. Methods Study cohort The parent cohort included 121 participants belonging to 34 households living, that had at least one member with a history of E. coli associated urinary tract infection or blood infection20. These index patients recovered and were discharged during the time of stool collection. All participants resided in Singapore during the period of sampling from 2017-201820. All participants consented to stool samples were screened for ST131 using culture-based, targeted PCR and whole genome sequences of isolates with epidemiological findings as previously reported20. The study flow diagram was based on previous reporting practice53. The samples utilized for shotgun metagenomic sequencing are a subset (42.5% or 255 samples) of the total human stool samples (n = 601) screened in the parent cohort20. Metadata for ST131 detection (positive or negative for ST131), host carrier status (persistent, intermittent, and non-carriers), age, sex (self-reported), antibiotics exposure for the duration of sampling, and household codes were used for analysis. Definition of host carrier status was based on the parent study, persistent carriers were defined as participants with at least two sequential samples positive for E. coli ST131, whereas intermittent carriers had no more than one sequential positive sample, and non-carriers had only ST131-negative samples20. The mean sampling frequency per subject was 2.14 ± 1.08 (s.d.). The mean interval between samples was 49.5 ± 34.7 (s.d) days for persistent carriers (n = 8 subjects), 53 ± 42.1 (s.d.) days for intermittent carriers (n = 28 subjects) and 64.2 ± 39.2 (s.d.) days for non-carriers (n = 53 subjects). There were 39 subjects with single timepoint samples (n = 1 for persistent carriers, n = 2 for intermittent carriers and n = 27 for non-carriers). The samples were further stratified by household risk to ascertain the microbial features associated with high-risk group (individuals from households with one persistent carrier and co-residents of any carrier status), low-risk (individuals residing with at least one intermittent carrier, in the absence of persistent carriers) and control (individuals from households with only non-carriers). The clinical isolates of six E. coli strains from six different index patients isolated from clinical samples at the point of hospitalization were randomly picked for whole-genome sequencing. Ethics approval was obtained from the National Healthcare Group Domain Specific Review Board (Reference: 2016/00998). Informed consent was sought from all individuals according to ICH Harmonised Good Clinical Practice (ICH E6 (R2)) guidelines as previously described20. DNA extraction, shotgun metagenomic and whole-genome sequencing Stool samples were stored at −80 oC prior to DNA extraction. DNA were extracted from 0.25 g wet-weight per stool sample and 2 mL E. coli overnight cultures using the manufacturers’ protocol in QIAamp PowerFecal Pro kit (Qiagen) and NucleoSpin Microbial Mini Kit (Machery-Nagel), respectively. For whole-genome sequencing (WGS), libraries were constructed using the NEBNext Ultra II library prep kit (New England Biolabs) according to the manufacturer’s protocol. WGS samples were paired-end (2 × 150 bp) sequenced on a NovaSeq 6000 sequencer (Illumina). Extracted DNA for metagenomic libraries were fragmented by Adaptive Focused Acoustics using a Covaris system. Metagenomic libraries for samples with MHH prefix (n = 244), including end repair, A-tailing, and adapter ligation, were performed with the GeneRead DNA Library I Core Kit (Qiagen) following the manufacturer’s instructions. Custom barcode adapters were substituted for the GeneRead Adapter I Set, and library amplification was carried out using custom index primers based on a protocol adapted from the Illumina multiplexing sample preparation oligonucleotide kit. For samples with non-MHH prefix sample identities (n = 31), metagenomic libraries were prepared according to the protocol in the NEBNext Ultra DNA library prep kit (New England Biolabs). All final libraries were assessed for quality and quantified using an Agilent Bioanalyzer with the DNA 1000 Kit (Agilent Technologies). Paired-end sequencing (2 × 101 bp) (MHH prefix samples) and 2 × 150 bp (non-MHH prefix samples) were performed on a HiSeq 4000 sequencer (Illumina) and NovaSeq 6000 sequencer (Illumina), respectively. Metagenomic analysis, plots, statistics and reproducibility Default parameters were applied to all in silico analysis unless stated otherwise. Metagenomic taxonomic and functional profiles were assigned using the bioBakery meta’omics workflow that includes KneadData v0.7.7-alpha for quality control and MetaPhlan 4 for taxonomic profiling and HuMAnN v3.0 for functional genes enzyme commission (EC; UniRef90) and MetaCyc pathway assignments54,55. Antibiotic resistance genes (ARG) profiles were obtained using ShortBRED v0.9.4 by mapping against CARD at the contig level56,57. All downstream analysis, and plots were generated using R v4.5.1 in Rstudio v2025.5.1 platform with enhanced features using tidyverse v2.0.058, finalfit v1.1.059, broom.mixed v0.2.9.660, reshape v0.8.1061, and patchwork v1.3.262. The R library vegan v2.7.1 and ape v5.8.1 were used for alpha- and beta-diversity analyses including PERMANOVA (ADONIS) test63,64. For the ST131-detection analysis, individual samples were treated as independent observations rather than using a longitudinal blocking factor (e.g., subject_code). This approach was adopted due to the highly variable sampling intervals and the limited number of longitudinal samples per subject. For sensitivity analysis, the ST131-negative group was randomly subsampled (n = 36 subjects) in a single pass to match the sample size of the ST131-positive group. Heatmaps were generated using ComplexHeatmap v0.12.565. Phylogenetic trees were constructed using ape v5.8.1 and taxize v0.10.0. Differential abundance analysis was performed using lmerTest v3.1-3 on log 10 transformed values for bacterial species, pathways (MetaCyc), and ARG relative abundances66. Microbial features with less than 5% prevalence across samples were excluded from differential abundance analysis. Pseudo-count is based on half of the minimal abundance of each feature. To assess multicollinearity among predictors, the Variance Inflation Factor (VIF) was calculated using the R car v3.1-5 package where generalized VIF (GVIF) adjusted for degrees of freedom showed low collinearity among covariates (GVIF < 2) (Suppl. Table 3)67. Reference groups for ST131 detection and ST131-carrier status were ST131-negative and non-carriers, respectively. The per-feature univariate regression models followed the general formula, feature ~ (intercept) + Predictor + (1|subject_code). Three separate models were tested using ST131 detection, carrier status alone, and household risk as primary predictors. In all cases, a random intercept for ‘subject_code’ was included to account for repeated measures or within-subject correlation. The formula for multivariate adjusted model was feature ~ (intercept) + ST131 detection + carrier status + sex + age + antibiotic usage + (1|subject_code). Statistically significant features (false-discovery rate (FDR) q-value < 0.05; log10-fold change > 0.3) were identified using the lmerTest package, utilizing analysis of variance test via Satterthwaite’s method for degrees of freedom66. Spearman correlation, Wilcoxon rank-sum, Kruskal-Wallis and Dunn’s posthoc tests and FDR adjusted P-values were performed using rstatix v0.7.268. EC genes and pathways abundance in reads per kilobase (RPK) were normalized to copies per million (CPM). For each analysis, no statistical method was used to predetermine sample size. No data were excluded from analysis. The experiments were not randomized. Randomization was not applicable to the analysis as participants were allocated to experimental groups based on distinct clinical phenotypes. Covariates used in the study were obtained directly from the primary study to ensure accurate baseline characteristics. The investigators were not blinded to group allocation during experiments and outcome assessment. However, to minimize bias, investigators were strictly blinded to subject identity during sample processing and downstream data analysis based on de-identified alpha-numeric codes. Whole genome sequencing analysis Raw paired-end reads were quality-filtered using Fastp v0.23.2 with changes to the following options (-u 0.5, -l 150, -overlap_diff_percent_limit 10)69. Filtered reads were assembled using SPAdes v3.15.5, and assembled scaffolds were polished using Polypolish v0.6.070,71. Protein-coding regions (CDS) and RNA genes were annotated using Prokka v1.14.572. The quality of the genomes was assessed using CheckM v1.2.273. The average nucleotide identity with BLAST (ANIb) was calculated using JSpeciesW (online accessed 16 Jan 2024)74. ST131 identity was confirmed using the MLST tool v2.0.9 database https://cge.food.dtu.dk/services/MLST/ that is based on the seven-locus scheme according to the Achtman typing method75–77. Serotyping was performed using the SerotypeFinder v.2.0.1 database v1.0.0 (accessed 16 Jan 2024) https://cge.food.dtu.dk/services/SerotypeFinder/78. ARGs were annotated from protein sequences using the Resistance Genes Identifier v6.0.3 tool against the Comprehensive Antibiotic Resistance Database (CARD) v3.2.8 online (accessed 16 Jan 2024) https://card.mcmaster.ca/57. Plasmids were determined from genomes using PlasmidFinder v2.0.1 (95% minimum identity; 60% minimum coverage) against the Enterobacteriales database last updated 18 Jan 2023, online tool (accessed 17 Jan 2024) https://cge.food.dtu.dk/services/PlasmidFinder/79,80. Siderophore proteins were identified by annotating each genome-derived amino acid sequences using BLASTKoala against the KEGG Ortholog (KO) database (v3.1 last updated 2025-01-01, accessed 01 Dec 2025)81. To identify the putative 1,5-anhydrousfructose reductase, amino acid sequences of the five ST131 genomes were searched against a Hidden Markov Model (HMM) trained KO database using KOfamKoala (v2026-04-01, KEGG release 118.0, accessed 10 May 2026)82. Potential homologs are proteins that exhibited the highest score-to-threshold ratio. Protein sequences for E. coli K12 substrain MG1655 (GCA_000005845.2) and C600 (GenBank CP031214.1) were downloaded from NCBI dataset83,84. MEGA v11 was used to align the protein sequences with MUSCLE and to build a maximum likelihood tree for the putative 1,5-anhydrousfructose proteins with reference protein sequences for EC 1.1.1.292 (UniProt Q2I8V6·AFR_ENSAD) and EC 1.1.1.263 (UniProt Q96JD6·AKCL2_HUMAN)85,86. The phylogenetic tree was based on a JTT matrix-model over 332 positions with bootstrap support inferred from 100 replicates87,88. Random forest regression model Random forest analyses were conducted using randomForest v47.1.2 (seed=123)89, caret v7.0.190, and pROC v1.19.0.1 packages91. Three random forest models were developed to classify persistent carriers and non-carriers or intermittent carriers: a clinical model using age, sex, and antibiotic usage; a pathway model using selected 15 pathway features, and a taxa model using selected 15 taxa features. Age was Z-score transformed, pathway and taxa abundances were log10-transformed (log10 abundance + 1 × 10⁻⁶) and Z-score transformed. Recursive feature elimination (RFE) was used to select the most discriminative features for pathway and taxa models. RFE was performed separately using a stepwise approach based on Gini impurity scores, progressively reducing total features: 548 pathways to 15 pathways and 1930 microbial species to 15 species. All models were trained using subject-level grouped 5-fold cross-validation repeated 5 times to prevent data leakage between samples from the same individual. Models were built with 1,000 trees, and the mtry hyperparameter was optimized via grid search within each cross-validation fold. Model performance was evaluated using AUC-ROC with 95% confidence intervals, and pairwise model comparison was performed using DeLong’s test for correlated ROC curves. Feature importance was assessed using the mean decrease in Gini impurity from the model trained on all 548 MetaCyc pathways (for pathway analysis) and the model trained on all 1930 microbial species (for taxa-based analysis). Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Supplementary Information Reporting Summary Transparent Peer Review file