Distinct signatures in the human gut and oral microbiomes of gastric cancer

Summary

Microbiome dysbiosis is increasingly recognized as a hallmark of gastric cancer (GC). Here, we analyzed gut and oral shotgun metagenomic data from 317 individuals across two independent cohorts, with validation in a Harbin cohort. We identify 20 oral-gut shared species enriched in the gut of GC, predominantly lactic acid bacteria (LAB). While most gut microbial markers are abundant in saliva, none are significantly altered in GC. Strain-level analysis of 87 matched saliva-stool metagenomes confirms oral-gut transmission of Streptococcus species. GC-enriched LAB form robust co-abundance networks in oral and gut microbiomes, suggesting synergistic interactions. Functional analysis reveals enriched lactate fermentation pathways in GC stool, aligning with LAB dominance and previous findings on gastric microbiota. Moreover, microbiome-based classifiers achieve high predictive accuracy (area under receiver operating characteristic curve [AUROC] = 0.85 for stool, 0.87 for saliva) for GC diagnosis, highlighting translational potential. Collectively, these findings underscore the critical role of the oral-gut microbiome axis in GC.

Graphical abstract

Highlights

• •Oral-gut shared species enriched in the gut microbiomes of GC
• •Streptococcus species transmission from mouth to gut confirmed at strain level
• •Microbiome classifiers enable accurate noninvasive GC detection

Qin et al. reveal distinct but coordinated oral-gut microbiome signatures in GC. Oral-gut shared lactic acid bacteria and enriched lactate fermentation pathways highlight a potential microbial axis in disease and support microbiome-based approaches for noninvasive detection.

Introduction

Gastric cancer (GC) ranks as the fifth most common cancer worldwide, both in incidence and cancer-related mortality.¹ East Asia bears a disproportionate burden, accounting for more than half of GC related deaths. Helicobacter pylori (Hp) infection is a well-established risk factor for GC, affecting more than half of the global population.² However, fewer than 5% of Hp-infected individuals develop GC, and Hp eradication only modestly reduces GC incidence in randomized controlled trials.³ This suggests that additional microbial factors contribute to GC development, necessitating a broader investigation of microbiome in GC pathogenesis.

Emerging evidence highlights non-Hp microbes in GC. High-throughput sequencing studies consistently report enrichment of Streptococcus and Lactobacillus—lactic acid bacteria (LAB)⁴—in the gastric microbiome of GC patients, particularly in tumors compared to adjacent tissues.^5,6,7,8,9,10 LAB metabolites, such as lactate, may promote carcinogenesis by altering the gastric microenvironment.¹¹ LAB was found to be associated with therapeutic response in multiple cancer types,¹² emphasizing their important roles in cancer biology. Recently, a mouse study showed that Streptococcus anginosus (Sa) promoted GC through interactions between its surface protein TMPC and gastric epithelial cells,¹³ confirming the oncogenic potential of non-Hp bacteria. In GC patients, intra-tumor Streptococcus was linked with worse prognosis.¹⁴ These gastric microbiome alterations suggest a complex microbial ecosystem influencing GC.

Our previous work demonstrated elevated Sa abundance in fecal samples from GC patients compared to gastritis patients,¹⁵ suggesting gut microbiota alterations mirror gastric dysbiosis. Several studies using 16S rRNA gene sequencing have identified variations in gut microbiota associated with GC, including enrichment of oral commensals such as Streptococcus at the genus level.^{16,17,18,19,20,21,22} However, the limited resolution of 16S rRNA sequencing precludes species-level insights critical for mechanistic understanding. A small-scale shotgun metagenomic study (32 GC and 33 chronic gastritis [ChG]) found enriched Bacteroides caccae and Sa in GC gut microbiomes but lacked validation.²³ Recent evidence also suggests distinct oral microbiomes in GC patients, with potential for risk assessment.²⁴ Given the mouth-stomach-gut anatomical connectivity, oral microbes may transmit to the gut, being a dysbiosis marker linked to diseases like colorectal cancer (CRC).²⁵ While strain-level analysis of the CRC pathogen Fusobacterium nucleatum showed identical oral and gut strains with differing dominance,²⁶ no study has investigated the strain-level similarity between the oral and gut microbes in GC.

To address the gaps mentioned above, we conducted a comprehensive shotgun metagenomic analysis of gut and oral microbiota in two independent cohorts, comprising 317 individuals—the largest such cohort to date, with validation in a third cohort from Harbin. We identified 20 oral-gut shared species, including LAB (Streptococcus spp. and Lactobacillus spp.), enriched in GC gut microbiome but not in oral microbiome. We further demonstrated enrichment of lactate fermentation pathways in GC stool, aligning with LAB dominance and prior gastric microbiota findings.^5,6,7,8,9,10 Using matched stool-saliva metagenomes from 87 patients, we confirmed oral-gut transmission of GC-enriched Streptococcus species via strain-level analysis. In addition, machine learning models based on our microbial signatures achieved high predictive accuracy. Our results validate previously reported microbial markers, identify signatures associated with GC, and provide evidence for the oral origins of GC-associated bacteria.

Results

Shotgun metagenomics sequencing of 404 stool and saliva samples

To study microbial markers associated with GC, we carried out shotgun metagenomics for stool and saliva samples from patients with GC and ChG (Figure 1A; STAR Methods). Stool metagenomes were generated from two independent cohorts: Cohort 1 included 106 ChG and 106 GC patients and Cohort 2 included 52 ChG and 53 GC patients. Demographic characteristics, including gender, age, and body mass index (BMI), were comparable between cohorts (Figures S1A and S1B).

Figure 1.

Overall characteristics of 404 stool and saliva metagenomes

(A) Study design.

(B) Principal coordinate analysis (PCoA) plot based on Bray-Curtis distance of species-level abundance profiles for 404 samples (312 stool and 92 saliva). Stool samples (Cohort 1 and Cohort 2) are colored in brown and saliva samples (Cohort 2) in green. Each point represents one sample.

(C) Boxplots of alpha diversity (Shannon index and observed species) for stool and saliva metagenomes across Cohort 1 and Cohort 2.

(D) PCoA of 212 stool metagenomes from Cohort 1.

(E) PCoA of 100 stool metagenomes from Cohort 2.

(F) PCoA of 92 saliva metagenomes from Cohort 2.

In (D–F), samples from GC and ChG patients are colored red and blue, respectively. Tables below (D–F) show permutational multivariate analysis of variance (PERMANOVA) results for GC status, age, gender, and body mass index (BMI). Significance (Wilcoxon rank-sum test): ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001, ns p ≥ 0.05.

We generated 312 stool metagenomes, yielding an average of 11.7 Gb (Cohort 1) and 12.2 Gb (Cohort 2) of high-quality data per sample after quality control (Table S1). For Cohort 2, deep shotgun metagenomic sequencing (>20 Gb raw data per sample) was applied to 92 saliva samples, resulting in an average of 15.0 Gb of high-quality, human-DNA-free data per sample (Table S1). Thus, the study produced three datasets: stool metagenomes from Cohort 1 (Stool 1), stool metagenomes from Cohort 2 (Stool 2), and saliva metagenomes from Cohort 2 (Saliva 2). Notably, 87 patients (43 ChG and 44 GC) had paired stool and saliva metagenomes.

Principal coordinate analysis (PCoA) revealed clear separation between stool and saliva microbiomes, with no distinct clustering between stool samples from Cohorts 1 and 2 (Figure 1B). Saliva microbiomes exhibited higher alpha diversity, measured by Shannon index and observed species, compared with stool microbiomes (Figure 1C). Beta diversity analysis showed greater differences in microbial composition between GC and ChG patients for saliva microbiomes than for stool microbiomes (Figures 1D–1F). Permutational multivariate analysis of variance (PERMANOVA) indicated that cancer status explained a larger proportion of variance in both stool and saliva microbiomes than age, gender, or BMI. After adjusting for age, gender, and BMI, no significant differences in alpha diversity were observed between GC and ChG patients for either gut or oral microbiomes (Figures S1C–S1H).

Microbial species differentiated between ChG and GC in the gut and oral cavity

To identify gut microbial species associated with GC, we performed a two-step validation metagenome-wide association study (MWAS) of species abundance in Cohort 1 and 2 using shotgun metagenomic data (STAR Methods). First, we tested the associations between abundances of 485 species (presented in >10% samples) and GC status in Cohort 1, using three methods, Wilcoxon rank-sum test, MaAsLin2,²⁷ and ANCOM-BC.²⁸ At a false discovery rate (FDR)-adjusted p < 0.05, 60 species were significantly differentiated between GC and ChG patients (Figure 2A). Of these, 28 species were validated in Cohort 2 at nominal p < 0.05 (Table S2; Figure 2B). We further investigated the potential confounding effects of age, gender, and BMI on the 28 differential species using MaAsLin2. Notably, 24 species remained significant post-adjustment (Table S2); the three of four non-significant species were only detected in 36–73 individuals, likely due to reduced power in the multivariate model. A sensitivity analysis on the Hp treatment-naive samples (95 GC vs. 92 ChG) showed that 27/28 species remained significantly differential. Subsequent analyses thus focused on these 28 validated gut species markers.

Figure 2.

Microbial species differentially abundant between GC and ChG in gut and oral cavity

(A) Two-step validation framework for species-level differential analysis in stool samples. Full testing results were provided in Table S2.

(B) Heatmap of 28 species differentially abundant between GC and ChG in 212 stool metagenomes from Cohort 1 (FDR-adjusted p < 0.05 in Cohort 1, unadjusted p < 0.05 in Cohort 2). Colored bars on the left indicate prevalence in Cohort 1. Species names are colored blue (ChG-enriched) or red (GC-enriched). Bold species have higher prevalence in saliva than stool in Cohort 2. Black asterisks denote oral-gut transmitters per Schmidt et al. (2019).

(C) Statistical framework for species-level differential analysis in oral samples. Full testing results were provided in Table S3.

(D) Heatmap of 36 species differentially abundant between GC and ChG in 92 saliva metagenomes from Cohort 2 (FDR-adjusted p < 0.05). Eighteen species statistically validated in the Harbin tongue samples were highlighted in bold. Colored bars on the left indicate prevalence in Cohort 2.

Among the 28 differential gut species, 23 were enriched in GC patients, while 5 were depleted compared to ChG patients (Figures 2B; S2). The GC-depleted species included beneficial bacteria such as Bacteroides ovatus and Fusicatenibacter saccharivorans. Notably, most GC-enriched species were oral commensals, with 11 previously reported as oral-gut transmitters.²⁹ Eight species belonged to the Streptococcus genus, including Sa. The association between Sa and GC was the strongest. To determine whether associations of other Streptococcus species with GC were independent of Sa, we conducted a conditional analysis (STAR Methods). Five species—S. constellatus, S. gordonii, S. infantis, S. mitis, and S. oralis—remained significantly associated with GC independent of Sa (Table S2), highlighting the prominence of Streptococcus species in the GC gut microbiome. Moreover, age is not a significant variable in the MaAsLin2 models (Table S2), indicating that Streptococcus enrichment is primarily a GC hallmark rather than an age effect. Other GC-enriched species included LAB,³⁰ such as Ligilactobacillus salivarius, Limosilactobacillus fermentum, Bifidobacterium dentium, Rothia SGB49305 (species-level genome bin³¹), Lactobacillus gasseri, Limosilactobacillus oris, and Limosilactobacillus vaginalis. Limosilactobacillus and Ligilactobacillus are recently reclassified genera from the former Lactobacillus genus based on genomic similarity.³² As Streptococcus species are LAB,⁴ these findings align with prior reports of LAB enrichment in GC gastric microbiota.¹¹

For the saliva microbiome, we identified 36 bacterial species that differed between GC and ChG patients (FDR <0.05, Table S3; Figure 2C). The most significant association was the decreasing abundance of Prevotella nanceiensis in GC. The abundance of Sa was slightly higher in GC than ChG (p < 0.1). In contrast to the gut microbiome, most differential species in saliva (29/36) were depleted in GC patients, with only seven species enriched (Figure 2D). Notably, 22 of 36 differential species in saliva and 10/28 in stool were present in at least 50% of patients, suggesting that GC-associated changes in the saliva microbiome involve more common species, whereas gut microbiome changes involve rarer species. We validated these findings in the in-house Harbin cohort (57 tongue metagenomes: 21 GC, 36 ChG; STAR Methods). Despite inherent differences between saliva and tongue microbiomes, 18/36 differential species were statistically validated (p < 0.05) and 30 showed directionally consistent changes (Table S3).

We next assessed the predictive utility of these bacterial markers for GC status (STAR Methods). For the 28 differential stool species, Boruta³³ identified 17 as more relevant than random probes (Figure 3A). These included 10 oral-gut transmitters, 7 of which ranked highest in importance, emphasizing their disease relevance. Classifiers based on these 17 markers, evaluated across six methods, yielded area under the receiver operating characteristic curve (AUROC) of 0.76–0.85 in the testing data, with random forest performing best (Figures S3A and S3B). The random forest model trained on Cohort 1 stool samples predicted Cohort 2 stool samples with an AUROC of 0.85 (95% confidence interval [CI]: 0.78–0.93; Figure 3B) and accuracy of 0.80 (95% CI: 0.71–0.87; Figure 3C). For the 36 differential saliva species, 20 exceeded random probe relevance (Figure 3D). Random forest again excelled in training and testing (Figures S3C and S3D). The model trained on Cohort 2 saliva samples predicted Harbin tongue samples with an AUROC of 0.87 (95% CI: 0.77–0.96; Figure 3E) and accuracy of 0.82 (95% CI: 0.70–0.91; Figure 3F). The saliva model included Corynebacterium matruchotii, which was identified as the most predictive species. The association between C. matruchotii abundance and GC was supported by Wilcoxon and ANCOM-BC (FDR <0.05), but not MaAsLin2 (Table S3).

Figure 3.

Predictive performance of gut and oral bacterial markers for GC

(A) Stool microbial species markers identified by the Boruta algorithm, ranked by median importance scores. Black asterisks denote oral-gut transmitters per Schmidt et al. (2019).

(B) Receiver operating characteristic curve (ROC) for the random forest classifier. The model was trained on Cohort 1 stool samples using the 17 markers from (A) and tested on Cohort 2 stool samples.

(C) Confusion matrix for the best classification in (B).

(D) Saliva microbial species markers identified by the Boruta algorithm, ranked by median importance scores.

(E) ROC for the random forest classifier. The model was trained on Cohort 2 saliva samples using 21 markers from (D) and tested in Harbin cohort tongue samples.

(F) Confusion matrix for the best classification in (E).

AUROC, area under receiver operating characteristic curve; Sen., sensitivity; Spe., specificity.

GC-enriched species in the gut were dominated by oral-gut shared species

To investigate oral-gut microbial signatures in GC, we analyzed shared microbial species using matched oral and gut metagenomes from Cohort 2. To minimize the influence of rare taxa, we filtered species with prevalence ≤5% in any dataset, retaining 1,163 species (Table S4). Of these, 206 species were present in all three datasets (Stool 1, Stool 2, Saliva 2; Figure 4A). Approximately half of these shared species belonged to the Firmicutes phylum, followed by Actinobacteria and Candidatus Saccharibacteria (Figure 4B). At the genus level, Streptococcus was the most represented (19 species), followed by Actinomyces (Figure 4B). We classified these 206 species as oral-gut shared and the remainder as non-shared. This classification was supported by strain-level analyses of Streptococcus spp. (below), which revealed significantly higher strain similarity in matched saliva-stool pairs than in inter-individual saliva metagenomes.

Figure 4.

Dominance of oral-gut shared species among GC-enriched gut microbiota

(A) Venn diagram showing the number of species detected in stool (Cohort 1 and Cohort 2) and saliva (Cohort 2) metagenomes. Only species detected in >5% individuals in stool or saliva samples of any cohort were counted.

(B) The phylum and genus compositions of 206 oral-gut shared species detected in all datasets. Bars are colored according to their belonging phylum. All phyla and the top 10 genera are shown.

(C) Boxplots of prevalence for oral-gut shared vs. non-shared species in stool and saliva metagenomes. Significance (Wilcoxon rank-sum test): ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001, ns p ≥ 0.05.

(D) Boxplots of summed abundance for oral-gut shared vs. non-shared species in stool and saliva metagenomes. Significance (Wilcoxon rank-sum test): ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001, ns p ≥ 0.05.

(E) Differentially abundant species stratified by oral-gut sharing status.

(F) Boxplots of prevalence for GC-enriched vs. non-enriched oral-gut shared species in stool. Significance (Wilcoxon rank-sum test): ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001, ns p ≥ 0.05.

Shared species exhibited distinct prevalence and abundance patterns in oral and gut microbiomes. In saliva, shared species had a median prevalence of 59.8%, significantly higher than non-shared species (median 40.2%; Figure 4C). In contrast, in the gut, most shared species were present in <12% of patients, with lower prevalence than non-shared species. Shared species contributed a median of 59.6% to total microbial abundance in saliva but only 6.4% in stool (Figure 4D). Notably, in the gut, shared species showed significantly higher prevalence in GC patients compared to ChG patients, whereas no such difference was observed in saliva (Figure S4).

We next categorized differential species (identified in the previous section) into shared and non-shared groups. In the saliva microbiome, only 7 of 36 differential species were shared, with most GC-depleted species belonging to the non-shared group (Figure 4E). In the gut microbiome, 20 of 28 differential species were oral-gut shared and all were GC enriched (Figure 4E). These 20 GC-enriched shared species had significantly higher prevalence than other shared species in the gut (Figure 4F). These findings highlight the prominence of oral-gut shared species in driving GC-associated gut microbiome alterations.

To test whether GC-enriched gut species originate from the oral cavity, we performed strain-level analysis using matched saliva and stool metagenomes. Given the low prevalence and abundance of shared species in the gut (Figure 5A), we focused on eight Streptococcus species with >30% prevalence in stool, previously identified as oral-gut transmitters,²⁹ and higher abundance in saliva than stool: S. anginosus, S. constellatus, S. cristatus, S. gordonii, S. infantis, S. mitis, S. oralis, and S. vestibularis (STAR Methods). Population average nucleotide identity (popANI),³⁴ a measure of strain similarity, was significantly higher between matched saliva and stool metagenomes from the same individual compared to saliva samples from different individuals (Figure 5B). Specifically, 42%–83% of matched comparisons yielded popANI >0.999, compared to <3% of inter-individual saliva comparisons. Among the 87 individuals with matched metagenomes, 44 (23 GC, 21 ChG) had detectable popANI for at least one species (Table S5), with no age differences versus the remainder (p = 0.95). Per-species correlations between age and popANI were weak (Figure 5C), suggesting that the oral-gut transmission was not confounded by aging. These results support the oral origin of these Streptococcus species in the gut of GC patients and validate their role as oral-gut transmitters.²⁹

Figure 5.

Genetic similarity of GC-enriched oral-gut shared species between stool and saliva microbiome

(A) Abundance and prevalence of 20 GC-enriched oral-gut shared species in Cohort 2 individuals with matched stool and saliva samples. GC-enriched oral-gut shared species was defined in Figure 4E.

(B) (Top) Boxplots of population average nucleotide identity (popANI) for eight GC-enriched oral-gut shared Streptococcus species comparing (1) matched stool and saliva metagenomes from the same individuals, (2) saliva metagenomes from different individuals, and (3) stool metagenomes from different individuals. (Bottom) Proportion of qualified comparisons with popANI >0.999. Parentheses denote numerators (comparisons with popANI >0.999) over denominators (total qualified comparisons) per category. Qualified comparison was defined as >10,000 compared bases across samples with genome breadth >10% for target species. Full details were provided in Table S5. Significance (Wilcoxon rank-sum test): ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001, ns p ≥ 0.05.

(C) Scatterplots of age versus popANI for the eight Streptococcus species, computed from matched saliva and stool metagenomes.

GC-enriched species in the gut exhibited strong co-abundance correlations in both oral and gut microbiomes

Microbial communities within the same niche often exhibit close interactions, and co-abundance networks can reveal biological insights.^35,36 To explore microbial interactions in GC, we constructed co-abundance networks for oral and gut microbiomes using FastSpar³⁷ with a significance threshold of FDR-adjusted p < 0.05, as applied in Wu et al.³⁵ (STAR Methods).

Oral-gut shared species displayed more frequent and stronger co-abundance correlations than non-shared species in both stool and saliva microbiomes (Figures 6A and S4C). For shared species, 20.9%, 7.3%, and 8.7% of possible interactions were significant in Stool 1, Stool 2, and Saliva 2 datasets, respectively, compared to 8.0%, 1.8%, and 3.5% for non-shared species. Among shared species, GC-enriched species exhibited significantly more frequent and stronger correlations than GC-depleted or non-differential shared species (Figures 6B and S4D). Specifically, 74.7%, 36.8%, and 18.9% of possible interactions among GC-enriched species were significant in Stool 1, Stool 2, and Saliva 2 datasets, respectively, compared to 20.5%, 7.1%, and 8.6% for other shared species.

Figure 6.

Strong co-abundance correlations among GC-enriched oral-gut shared species

(A) Bar plot showing the proportion of significant co-abundance correlations (FDR-adjusted p < 0.05, calculated by FastSpar) for oral-gut shared vs. non-shared species in Stool 1, Stool 2, and Saliva 2 datasets. Proportions are calculated as significant correlations divided by total species-species pairs. Oral-gut shared species are defined in Figure 4A.

(B) Bar plot showing the proportion of significant co-abundance correlations for GC-enriched vs. non-enriched oral-gut shared species in Stool 1, Stool 2, and Saliva 2 datasets. GC-enriched species are defined in Figure 4E.

(C) Co-abundance network among 14 GC-enriched oral-gut shared species in our four datasets, visualized using Cytoscape. All reported correlations are positive. Node sizes were scaled to the mean relative abundance (0.02%, 4.51%), node colors were mapped to prevalence, and edge width was mapped to the correlation coefficient (0.24, 0.61). Limosilactobacillus fermentum was undetectable in the Harbin cohort tongue dataset.

Focusing on robust co-abundance networks among GC-enriched species present in all three datasets of Cohorts 1 and 2, we identified 15 positive correlations involving 14 species (Figure 6C). A tightly connected network was observed for LAB, including Limosilactobacillus oris, Limosilactobacillus vaginalis, Limosilactobacillus fermentum, Lactobacillus gasseri, and Bifidobacterium dentium, which extended to S. gordonii, Ligilactobacillus salivarius, and Rothia SGB49305 (species-level genome bin³¹). Additionally, Lancefieldella parvula correlated with Actinomyces oris, Mogibacterium diversum, and Schaalia meyeri. Of note, 10/15 positive correlations were validated in the Harbin tongue data at p < 0.05 (Figure 6C).

Enrichment of lactate fermentation pathways in the gut microbiome of GC patients

Functional metagenomics analysis using HUMAnN 3³⁸ revealed distinct pathway profiles (Methods). PCoA showed clear clustering of stool and saliva samples, while no segregation between stool samples from Cohorts 1 and 2 (Figure S5A). In contrast to the species profile, saliva microbiomes had lower alpha diversity than stool microbiomes (Figure S5B). PERMANOVA confirmed cancer status as the strongest predictor of pathway variance, surpassing age, gender, or BMI (Figures S5C–S5E). ChG patients had higher Shannon index than GC patients, with or without covariate adjustment (Figure S6).

In stool, we applied the two-step validation MWAS framework to identify MetaCyc pathways differentially abundant between GC and ChG (STAR Methods). In the first step, we tested the associations of 493 pathways (presented in >10% samples) and GC status in Cohort 1 using three methods (Wilcoxon, MaAsLin2, and ANCOM-BC); 77 pathways were nominated at FDR <0.05, and 29 of them were validated at p < 0.05 in Cohort 2 (Table S6). The strongest GC association was found in CDP-diacylglycerol biosynthesis (PWY-5981). In saliva, 76 pathways were differential (FDR-adjusted p < 0.05 in Wilcoxon, MaAsLin2, and ANCOM-BC tests), with 68 enriched in ChG (Table S7). Nine pathways were concordantly enriched in ChG in both stool and saliva microbiomes, while only two (P125-PWY and PWY0-1298) were concordantly GC enriched. Validation in Harbin tongue metagenomes was modest: 38/76 (50%) showed directionally consistent changes with 18 (23.7%) statistically validated (p < 0.05), reflecting niche-specific differences.

Given the enrichment of LAB in GC, we analyzed lactate fermentation pathways in stool metagenomes. Homolactic (ANAEROFRUCAT-PWY) and heterolactic (P122-PWY) fermentation pathways were significantly enriched in Cohort 1 GC patients (FDR-adjusted p < 0.001, Table S6) and directionally consistent in Cohort 2. The lack of significance in Cohort 2 may be due to smaller sample size. The enrichment of pathway P122-PWY remained significant in the sensitivity analysis using a subset of the Hp treatment-naive samples (95 GC vs. 92 ChG, P_wilcoxon = 3.93 × 10⁻⁵, P_masslin = 1.28 × 10⁻³). Notably, seven enzymes involved in the hetero-fermentative pathway were all significantly enriched in Cohort 1 GC patients (Figure 7; Table S6). These included key enzymes that affect lactic acid production: phosphoketolase (4.1.2.9) and L- and D-lactate dehydrogenases (1.1.1.27 and 1.1.1.28).³⁹ For homo-fermentative pathway, seven of nine detected enzymes were significantly enriched in Cohort 1 GC patients (Figure S7; Table S6). These findings align with LAB enrichment, reinforcing their role in GC-associated gut dysbiosis.

Figure 7.

Enrichment of enzymes for heterolactic fermentation in the stool microbiome of GC patients

(A) Schematic graph of MetaCyc pathway P122-PWY: heterolactic fermentation. Key enzymes (noted as EC number) commonly found in lactic acid bacteria are indicated.

(B) Boxplots show the abundance of 7 enzymes in the stool microbiome of GC and ChG patients.

Significance (Wilcoxon rank-sum test on raw relative abundance): ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001, ns p ≥ 0.05. Statistical significance for enzymes 2.7.1.4, 1.1.1.49, 4.1.2.9, and 1.1.1.28 were supported by Wilcoxon test on center log-ratio-transformed values. Enzymes 2.7.1.4, 1.1.1.49, and 4.1.2.9 were supported by testing using MaAsLin2. Full results were provided in Table S6.

Discussion

Leveraging 404 metagenomes (312 stool and 92 saliva) from 317 individuals across two independent cohorts, with validation in a third cohort, we conducted a comprehensive MWAS of the gut and oral microbiomes in GC versus ChG. We identified 28 validated gut microbial markers, of which 20 were oral-gut shared species, predominantly LAB, including eight Streptococcus species and five Lactobacillus species, enriched in GC gut microbiome compared to ChG. While most gut microbial markers were abundant in saliva, none were differential between GC and ChG, underscoring distinct oral and gut signatures. Strain-level analyses of 87 matched saliva-stool pairs confirmed oral transmission for these eight Streptococcus species, with popANI >0.999 in 42%–83% of matched pairs versus <3% of inter-individual saliva comparisons. These findings extend prior gastric microbiome studies reporting Streptococcus and Lactobacillus enrichment in GC tissues,^5,6,7,8,9,10 highlighting a continuous “oral-gastric-gut” dysbiosis axis.

This LAB consortium’s oral origin raises a critical question: why do these oral commensals thrive ectopically in the GC gut? We propose that the GC host may provide a “permissive” microenvironment: the hypoacidic state associated with gastric tumors lowers the barrier to oral bacterial survival, the tumor microenvironment may offer nutritional sources supporting their growth, and the GC-associated local and systemic immunosuppression could impair the clearance of these “foreign” colonists. Thus, these oral bacteria may not be mere innocent “passers-by” but rather “opportunists” that are selectively favored and thrive within the ecosystem of GC. This “transmission to colonization” model frames non-Hp microbes as active contributors to GC progression. The GC-specific enrichment of multiple Streptococcus species distinguishes it from other cancers, such as CRC (where transmitters like Dialister pneumosintes, F. nucleatum, Gemella morbillorum, Parvimonas micra, Peptostreptococcus stomatis, Solobacterium moorei predominate),^40,41 and supports the oral cavity as a reservoir for gastrointestinal pathobionts.⁴²

Streptococcus species’ biofilm-forming capacity, driven by three quorum-sensing systems,⁴³ may enhance their ectopic colonization in the gut, explaining their prominence in GC. Specifically, S. gordonii recruits co-colonizers via multiple functional pathways to form polymicrobial biofilms,⁴⁴ potentially facilitating ectopic gut colonization in GC. Reciprocally, biofilms may enhance S. gordonii survival, as evidenced by its interactions with F. nucleatum that attenuate macrophage responses and promote mutual persistence.^45,46 In our study, Streptococcus and Lactobacillus species together with other LAB (B. dentium, Rothia spp., and Granulicatella adiacens [nutritionally variant Streptococcus]) formed a robust co-abundance network in both oral and gut microbiomes, suggesting that polymicrobial synergy may contribute to GC pathogenesis. This mirrors in vitro evidence of enhanced acid tolerance and biofilm synergy in LAB consortia⁴⁷ and mucosal Lactobacillus co-enrichments in GC.⁵ Such polymicrobial cooperation may sustain gastric persistence independently of Hp’s urease-mediated pH modulation, especially in Hp-negative cases, with dietary LAB sources potentially exacerbating oral-gut seeding.⁴⁸

Following their successful colonization, this consortium may foster a pro-tumorigenic milieu via metabolic outputs. Functional profiling revealed significant enrichment of homolactic and heterolactic fermentation pathways in GC stool, with all seven heterolactic enzymes (including phosphoketolase and lactate dehydrogenases)³⁹ and 6/9 homolactic enzymes elevated. This finding, while derived from stool samples, aligns with the dominant abundance of LAB and finds a compelling parallel in the gastric niche, where elevated lactate levels have been clinically detected in the gastric juice of GC patients.⁴⁹ It is thus plausible that lactate produced by these bacteria could contribute to acidifying the tumor microenvironment (pH ≤ 6.5), a condition known to activate matrix metalloproteinases that degrade the extracellular matrix and promote invasion.¹¹ Furthermore, lactate has been shown in other contexts to upregulate VEGFR2 expression, enhancing VEGF-mediated angiogenesis,⁵⁰ and to induce PD-1 expression in regulatory T cells, thereby contributing to treatment resistance.⁵¹ Our findings generate the hypothesis that gut LAB signatures may reflect broader digestive dysbiosis, with lactate as a potential mechanistic link to gastric oncogenesis. Direct validation of LAB signatures and lactic acid metabolism in gastric samples is required to test this hypothesis.

In models of gastric carcinogenesis, Hp acts as the primary initiator, whose chronic infection triggers a cascade of mucosal damage, inflammation, atrophy, and metaplasia.⁵² This process profoundly disrupts the gastric microenvironment and ecological barrier. Our findings extend this literature-derived initiator-promoter model as a hypothesis applicable to non-Hp LAB: Hp as initiator generates a permissive niche, while non-Hp LAB may act as promoters via biofilms, lactate metabolism, and immune modulation, thereby advancing carcinogenesis synergistically or independently. This may explain Hp-negative GC and residual risk following eradication and warrants direct testing in longitudinal studies and animal models.

Our microbial signatures are robust predictors (stool AUROC = 0.85, accuracy = 0.79; saliva = 0.87, 0.82), with oral models outperforming gut ones—suggesting earlier or more pronounced oral alterations. These hold promise for non-invasive GC risk stratification. Their utility in early risk predictions remains to be validated in prospective cohorts.

In conclusion, this study elucidates GC’s microbial landscape in the oral-gut axis, revealing distinct microbial signatures, LAB-dominated consortia, oral-gut transmission, and potential lactate-mediated mechanisms. By framing non-Hp opportunists in an initiator-promoter paradigm, our findings advance mechanistic understanding and nominate biomarkers for diagnostics and microbiome-targeted therapies in GC.

Limitations of the study

An important caveat of our study is that most analyzed samples were derived from Hp-negative patients. Our findings may not be generalizable to Hp-positive GC patients and should be interpreted within this specific clinical context. With widespread global adoption of Hp eradication therapies, the incidence of Hp-negative GC is increasing, rendering our findings particularly relevant to this emerging subtype. Second, we did not assess participants’ oral health status (e.g., caries prevalence), which could confound oral-gut transmission patterns. But most of the known caries-associated bacteria, such as Streptococcus spp., Bifidobacterium dentium, Actinomyces oris, and Granulicatella adiacens,⁵³ were not among the top hits in the differential abundance analysis of saliva metagenomes (Table S3), supporting GC specificity. Third, our shotgun metagenomic approach prioritized bacterial communities, as fungal and viral sequences were under-represented; targeted enrichment sequencing will be essential to interrogate their contributions to GC. Fourth, the modest saliva sample size (n = 92) and stringent statistical threshold (FDR < 0.05 in Wilcoxon, MaAsLin2, and ANCOM-BC) may have underestimated oral microbiome alterations and saliva may not fully capture the heterogeneity of oral niches (e.g., tongue, buccal mucosa, dental plaque). Nonetheless, saliva’s substantial daily flux into the gastrointestinal tract (0.5–1.5 L)⁵⁴ underscores its pertinence for probing oral-gut microbial relays in GC, supporting our focus on saliva for investigating oral-gut microbial transmission. Moreover, validation in the independent Harbin tongue cohort confirmed statistical significance for half of the GC-associated species and directional consistency for 30/36, alongside strong predictive performance (AUROC = 0.87), affirming the cross-niche robustness of our oral signatures. Last, the cross-sectional design precludes causal inference regarding whether oral-gut shared microbe enrichment drives or reflects GC progression. In addition, the stool microbiome is influenced by multiple factors throughout the entire gastrointestinal tract, liver, and pancreas. Prospective longitudinal studies are thus needed to dissect these dynamics and their mechanistic impacts.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Youwen Qin, PhD (qinyouwen@genomics.cn).

Materials availability

This study did not generate new reagents.

Data and code availability

• •Data: Raw reads of shotgun metagenomics sequencing data have been deposited into CNGB Sequence Archive (CNSA) of China National GeneBank Database (CNGBdb)⁵⁵ with accession number CNP0004342 (https://db.cngb.org/data_resources/project/CNP0004342/).
• •Code: The codes used in this study have been deposited at GitHub, https://github.com/Owen-haha/GC_OralGut_MWAS.
• •General statement: Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request.

Acknowledgments

This study was supported by the State Key R&D Program (2020YFA0509200), Noncommunicable Chronic Diseases-National Science and Technology Major Project (2025ZD0551704), National Natural Science Foundation of China (82530086, 82330086, 82203224, and 82570645), Shanghai Jiao Tong University 2030 Initiative (WH510272101), Heilongjiang Provincial Key Research and Development Program (no. 2023ZX07D05), Shanghai Rising-Star Program (24QA2705000), Shanghai Eastern Talent Plan Youth Project (QNWS2025114), Chenguang Program of Shanghai Education Development Foundation and Shanghai Municipal Education Commission (22CGA17), and Shanghai Youth Medical Talents – Specialist Program (2023-62). We thank our colleagues at BGI Genomics, specifically Dr. Yuanqiang Zou and Xin Tong, for their fruitful discussion and Qianhui Feng, Dr. Meng Ni, Yinbin Qiu, Zhun Shi, and Dr. Huahui Ren for their supports in bioinformatics. We thank Dr. Jia Li, Dr. Zi-Quan Sun, and Ze-Wen Chang for their contributions to the Harbin Cohort. We also thank our colleagues at Renji Hospital, Shanghai. We acknowledge the support from the “Harbin Large-Scale Population Screening Program for Digestive System Cancers and Cardiovascular Diseases related “Four High Risk” Factors” for providing part of the metagenomic data.

Author contributions

Y.Q., J.-Y.F., and C.-B.Z., conceived the project. Y.Q. designed the study, analyzed data, and drafted the manuscript. Y.-X.Z. designed the study, analyzed the data, and wrote the manuscript. L.-P.L. analyzed the data and wrote the manuscript. Y.-H.X., X.-Y.M., Y.H., L.-C.Z., and H.Z. revised the paper. J.-J.D., Y.H., K.S., and M.L. contributed to the validation. S.Z. and J.-Y.F. supervised the study, provided resources, sponsored funding, and reviewed and revised the paper.

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used ChatGPT, Grok, Gemini, and DeepSeek in order to polish the English writing. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Biological samples
Human saliva	Renji Hospital, Shanghai, China	N/A
Human stool	Renji Hospital, Shanghai, China	N/A
Human tongue	The Fourth Affiliated Hospital of Harbin Medical University, Harbin, China	N/A
Critical commercial assays
FastPure Stool DNA Isolation Kit (Magnetic bead)	MJYH, Shanghai, China	Cat# T10-100
FastPure Soil DNA Isolation Kit (Magnetic bead)	MJYH, Shanghai, China	Cat# T09-96
MagPure Stool DNA Kit (Magnetic bead)	Magen Biotech	Cat# MD5115-02B
Deposited data
Shotgun metagenomic sequencing data	This study	[CNSA]: [CNP0004342]
Software and algorithms
R	CRAN	https://cran-archive.r-project.org/
MetaPhlAn4	Blanco-Míguez et al.31	https://huttenhower.sph.harvard.edu/metaphlan/
HUMAnN3	Beghini et al.38	https://huttenhower.sph.harvard.edu/humann/
inStrain	Olm et al.34	https://instrain.readthedocs.io/en/latest/
MEGAHIT	Li et al.56	https://github.com/voutcn/megahit
MetaBAT	Kang et al.57	https://github.com/bioboxes/metaBAT
dRep	Olm et al.58	https://github.com/MrOlm/drep
checkM2	Chklovski et al.59	https://github.com/chklovski/CheckM2
GTDB-Tk	Chaumeil et al.60	https://github.com/Ecogenomics/GTDBTk
metapi	GitLab	https://github.com/ohmeta/metapi
FastSpar	Watts et al.37	https://github.com/scwatts/fastspar
fastp	Chen et al.61	https://github.com/OpenGene/fastp
Bowtie2	Langmead et al.62	https://github.com/BenLangmead/bowtie2
inkScape	GitLab	https://github.com/inkscape/inkscape
Cytoscape	Shannon et al.63	https://github.com/cytoscape/cytoscape

Experimental model and study participant details

Study cohorts and sample collection

Participants were recruited at Shanghai Renji Hospital after reading and signing informed consent and completing questionnaires. The study was approved by the Ethics Review Committee of Renji Hospital (LY2024-276-B). Cohort 1 participants (n = 212) were derived from our previous study (age 30–86, 117 male, Figure S1A),¹⁵ and Cohort 2 participants (n = 105) were recruited in 2024 (age 26–83, 54 male, Figure S1B). All participants underwent diagnostic gastroscopy to confirm GC or ChG. All ChG participants with a normal immunity were healthy, who showed no evidence of ulcers, polyps, tumors, severe erosions, moderate to severe bile reflux, or other significant abnormalities on gastroscopy performed within the past six months. Additionally, all participants underwent coloscopy within the past five years with no abnormalities detected.

Hp status and treatment history

After reviewing surgical pathology, rapid urase test (RUT), and breath test results, we confirmed that 97/106 (92%) of Cohort 1 GC patients were Hp-negative. This might be attributed to prior eradication therapy or the declining Hp infection rate observed in advanced GC stages. All Cohort 1 ChG participants were Hp-negative because active Hp infection typically causes significant gastric inflammation that would have excluded them from the study according to our strict enrollment criteria. In Cohort 2, ChG participants were Hp-negative based on RUT of gastric mucosa, and most of the GC patients were Hp-negative by rapid urease test or ¹³C urea breath test. None of the Hp-negative participants in our study received eradication therapy within the five years prior to enrollment. Patients who had received two or more eradication treatments had been excluded for Cohort 1 and 2. In Cohort1, 95/106 (88.5%) of GC patients and 92/106 (86.8%) of ChG patients were entirely treatment naive; the other small subset (11.5% and 13.2%, respectively) had received a single eradication course more than five years ago. In Cohort2, 100% of ChG participants were pure Hp-negative and completely treatment-naïve; only 6/50 of GC patients had a single treatment history over five years prior.

Stool samples were collected from both cohorts using provided kits, with instructions for in-hospital or at-home collection, before gastroscopy. Saliva samples were collected from Cohort 2 participants, who were instructed to refrain from eating, drinking and mouth cleaning for 30 min prior to sampling. Stool and saliva samples were immediately stored at −80°C before being shipped for sequencing.

For independent validation of oral metagenomic associations, we included an in-house cohort recruited at The Fourth Affiliated Hospital of Harbin Medical University (Harbin cohort). The study was approved by the Ethics Review Committee of The Fourth Affiliated Hospital of Harbin Medical University (HYDSYEX-2024-105) and was conducted in accordance with the relevant ethical and human genetic resource regulations. The Harbin cohort included 21 GC patients (age 37–83, 14 male) and 36 ChG participants (age 45–63, 12 male), following the same definition with Cohort 1 & 2. ChG participants had undergone colonoscopy with no abnormalities. Tongue swabs were collected from each participant by registered nurses. All participants signed informed consent.

Method details

Metagenomic sequencing experiments

DNA extraction

Total genomic DNA was extracted from stool samples using the FastPure Stool DNA Isolation Kit (Magnetic Bead) (MJYH, Shanghai, China) following the manufacturer’s instructions. For saliva samples, DNA was extracted using FastPure Soil DNA Isolation Kit (Magnetic bead, MJYH, Shanghai, China). DNA concentration and purity were measured using a Synergy HTX microplate reader and NanoDrop 2000 spectrophotometer (Thermo Fisher Scientific, Waltham, MA, USA), respectively. DNA quality was assessed via 1% agarose gel electrophoresis.

Library preparation and sequencing

DNA was fragmented to ∼400 bp using a Covaris M220 (Gene Company Limited, China). Paired-end libraries were constructed with the NEXTFLEX Rapid DNA-Seq Kit (Bioo Scientific, Austin, TX, USA) and sequenced on an Illumina NovaSeq X Plus platform (Illumina Inc., San Diego, CA, USA) using the NovaSeq X Series 25B Reagent Kit per manufacturer instructions. Targeted sequencing depths were 10 Gb for stool and 20 Gb for saliva samples.

Harbin cohort

DNA was extracted from tongue swabs using the MagPure stool DNA KF kit B (no. MD5115-02B). DNA concentrations were estimated using Qubit (Invitrogen). The DNA libraries were constructed from 200 ng input DNA. Shotgun metagenomic sequencing was then performed on the BGI-SEQ platform⁶⁴ to generate at least 20 million paired-end reads (100 bp length, >20Gb raw data) for each sample.

Metagenomic sequencing data analysis

Quality control

Sequencing data were processed using the metapi workflow (https://github.com/ohmeta/metapi). Reads with a median quality score <Q30 (0.001% error rate) or length <70 bp were removed, and adapter sequences and low-quality tails were trimmed using fastp.⁶¹ Remaining reads were mapped to the human reference genome (hg38) using Bowtie2⁶² with the “--very-sensitive” parameter to remove host DNA. High-quality, non-human reads (termed “useful reads”) were used for downstream analyses. Quality control metrics are summarized in Table S1.

Taxonomic and functional profiling

Species composition was determined using MetaPhlAn 4 with the mpa_vJun23_CHOCOPhlAnSGB_202403 database.³¹ Functional profiles were generated with HUMAnN 3.9³⁸ based on UniProt protein sequence reference clusters,⁶⁵ and used for a comprehensive analysis on corresponding functional grouping according to MetaCyc Pathways,⁶⁶ Enzyme Commission (EC) terms.

De novo assembly and MAG annotation

De novo assembly of 404 metagenomes (312 stool, 92 saliva) was performed using the metapi workflow. MEGAHIT v1.2.9⁵⁶ was used for assembly, followed by alignment with Bowtie2 and binning with MetaBAT2.⁵⁷ This resulting bins with a minimum length of 200 kb were defined as metagenome-assembled genomes (MAGs). MAGs quality was evaluated by CheckM2,⁵⁹ which found 3,677 (2,030 for Stool 1, 1,086 for Stool 2, and 561 for Saliva 2) high-quality MAGs per MISAG criteria (>90% completeness and <5% contamination).⁶⁷ MAGs were annotated using the GTDB toolkit v2.1.0⁶⁰ with GTDB release 220⁶⁸ and dereplicated using dRep v3.5.0.⁵⁸

Strain-level analysis

Due to low abundance of oral-gut shared species, no high-quality MAGs could be reconstructed for these taxa. Reference genomes for GC-enriched oral-gut shared species (Streptococcus spp., Actinomyces oris, Bifidobacterium dentium, Granulicatella adiacens, Lactobacillus gasseri, Lancefieldella parvula, Mogibacterium diversum, Schaalia meyeri) were downloaded from the National Center for Biotechnology Information (NCBI) genome database and dereplicated using dRep. Combined high-quality MAGs (n = 735) and reference genomes (n = 119) were analyzed with inStrain³⁴ to calculate genome breadth (fraction of genome mapped) and coverage (average mapping depth). Reference genomes for eight Streptococcus species included S. anginosus (NCBI: GCF_900636475.1), S. constellatus (NCBI: GCF_023167545.1), S. cristatus (NCBI: GCF_000222765.1), S. gordonii (NCBI: GCF_001553855.1), S. infantis (NCBI: GCF_000187465.1), S. mitis (NCBI: GCF_000148585.2), S. oralis (NCBI: GCF_900637025.1), and S. vestibularis (NCBI: GCF_000188295.1). Strain-level pairwise average nucleotide identity (popANI) was then calculated between (1) matched stool and saliva metagenomes from the same individuals, (2) interl-individual saliva metagenomes and (3) inter-individual stool metagenomes in Cohort 2. Only comparisons exceeding 10,000 aligned bases across samples with genome breadth >10% for the target species were retained for downstream analysis. At this threshold, inferences were limited to the eight Streptococcus species.

Quantification and statistical analysis

Microbial diversity was assessed using species profiles from MetaPhlAn 4. Alpha diversity was calculated as Shannon index and observed species number, and beta diversity was measured using Bray-Curtis distance. We used permutation multivariate analysis of variance (PERMANOVA) to estimate the variance explained by GC, age, gender (self-report), body mass index (BMI). For each test, the number of permutations was set to 100,000, which is capped at P ≤ 1E−5. R package vegan (version 2.6-2) was used for these calculations.

Differential abundant analysis

We performed metagenome-wide association (MWAS) on species and pathway levels using three methods: two-sided Wilcoxon-Mann-Whitney test (also named as Wilcoxon rank-sum test), MaAsLin2 (v1.16.0) and ANCOM-BC. Both MaAsLin2 and ANCOM-BC are designed for compositional data. ANCOM-BC also embeds a sensitivity analysis for pseudo-count addition. Benjamini-Hochberg FDR method was used for multiple testing correction.

For stool microbiomes, a two-step framework was applied. In step1 (based on Cohort1), significant markers required to be: (1) FDR-adjusted p < 0.05 in Wilcoxon, MaAsLin2 and ANCOM-BC tests and (2) passing the pseudo-count sensitivity analysis in ANCOM-BC. In step2, the candidate markers from step1 were further tested in Cohort 2 using Wilcoxon and MaAsLin2. ANCOM-BC was not used in the subset analysis as the embedded bias-correction is benchmarked on entire metagenome. Markers with nominal p < 0.05 in Wilcoxon or MaAsLin2 were counted into the final set. The potential confounding effects of age, gender, and BMI was tested by including them as covariates in the MaAsLin2 models.

For saliva microbiomes, differential markers were defined as FDR <0.05 in the three methods (Wilcoxon, MaAsLin2 and ANCOM-BC). Despite we validated the saliva markers in the Harbin tongue samples, we did not consider it as a second validation step as the inherit niche difference between saliva and tongue samples.

The input for Wilcoxon and MaAsLin2 was center log-ratio (CLR)-transformed relative abundances,⁶⁹ calculated by R package compositions (v2.0-7) with zeros were imputed as 1/10 of the minimum non-zero value. MaAsLin2 parameters normalization and transform were set as “NONE”. The input data for ANCOM-BC was sequence count table from MetaPhlAn4. The MWAS was applied to species and pathways presented in >10% samples.

Conditional and sensitivity analysis

To investigate the potential confounding effect of Sa on 7 non-Sa Streptococcus species, we performed conditional analysis using MaAsLin2. For each non-Sa Streptococcus species, we tested the association with GC status by adding Sa abundance as a covariate. To investigate the potential impact of Hp treatment, a sensitivity analysis was run on the treatment naive subset for identified markers, using Wilcoxon and MaAsLin2.

Disease classification

Feature selection for machine learning was conducted using Boruta (p-value = 0.001, Monte Carlo adjustment = TRUE, maxRuns = 200, seed = 11; R package v9.0.0),³³ an all-relevant wrapper algorithm that iteratively eliminates features less informative than permuted shadow variables. We evaluated six interpretable machine learning methods which are commonly applied in biological data: support vector machine with linear kernel (SVM-LIN), random forest (RF), linear discriminant analysis (LDA), generalized linear model (GLM), neural network (NNet), and generalized boosted model (GBM). Models were trained and tuned using the caret package (v7.0-1) with 10-fold cross-validation. Performance was quantified by area under the receiver operating characteristic curve (AUROC) via the pROC package (1.19.0.1), with the final model selected based on highest AUROC in the external validation dataset.

Co-abundance correlations were calculated using FastSpar (v1.0.0)³⁷ with 1,000 times bootstrap (FDR-adjusted p < 0.05 for significance), following Wu et al.³⁵ Correlation networks were visualized using Cytoscape (v3.10.3),⁶³ using median bootstrap correlation strengths.

Published: April 20, 2026