Cross-site oral and esophageal microbiome signatures define diagnostic patterns and reveal mechanistic drivers of ESCC

Summary

The oral cavity and esophagus are contiguous mucosal sites whose microbiomes may jointly influence esophageal squamous cell carcinoma (ESCC). To address whether microbial alterations are shared across oral and esophageal niches and possess diagnostic potential in ESCC, we performed five-region 16S rRNA sequencing on paired oral swabs and esophageal tissues from 45 patients with ESCC and matched controls. We identified consistent enrichment of Porphyromonas, Fusobacterium, and Treponema, with the depletion of Neisseria, Rothia, and Actinomyces across both sites. These cross-site signatures showed strong diagnostic performance, supporting proof-of-concept non-invasive ESCC prediction using oral swabs. Functional prediction suggested altered microbiome-associated functional profiles, including enrichment of glycan- and amino acid-related pathways and reduced fatty acid metabolism. In a 4-NQO mouse model, P. gingivalis accelerated ESCC development and promoted inflammatory and immune-suppressive responses. Together, these findings identify shared oral-esophageal microbial signatures with potential diagnostic value in ESCC and support further validation of oral microbiome-based detection strategies.

Graphical abstract

Highlights

• •Cross-site oral-esophageal microbial signatures in ESCC feature enriched anaerobes
• •Microbiome signatures support ESCC detection and patient stratification
• •ESCC microbiomes exhibit predicted alterations in epithelial and metabolic pathways
• •P. gingivalis promotes ESCC progression and reshapes tumor immunity

Biological sciences; Microbiome; Cancer

Introduction

The human orodigestive tract harbors a complex and dynamic microbiome that plays pivotal roles in maintaining mucosal homeostasis, regulating immune responses, and influencing disease susceptibility.^1,2 Disruption of microbial community composition, termed microbial dysbiosis, has emerged as a key driver of inflammation-associated carcinogenesis across multiple anatomical sites.^3,4 High-throughput sequencing studies have revealed enriched pathobionts and depleted commensals in the tumor microenvironment of the gastrointestinal tract (GI) and orodigestive malignancies, underscoring the systemic impact of the microbiome on cancer risk, progression, and therapeutic response.^5,6 In several cancers, specific microbial taxa have been implicated in tumor biology through both direct and immune-mediated mechanisms. For example, Fusobacterium nucleatum promotes cancer progression via adhesin-mediated cancer cell invasion and immune suppression,^7,8 whereas oral pathogens such as Porphyromonas and Tannerella contribute to extracellular matrix degradation and tumor invasiveness.^9,10 Conversely, commensal Streptococcus species such as Streptococcus cristatus may exert protective effects, and their reduction or depletion exacerbates microbial imbalance.^11,12 Collectively, these findings illustrate how tumor-associated microbial shifts often involve both enrichment of oncogenic species with loss of protective taxa, thereby fostering a pro-tumorigenic microenvironment.

Oral pathogens represent a clinically important subset of tumor-associated microbiota that may contribute to tumor progression through persistent mucosa colonization, translocation to distal sites, and modulation of host immunity.^13,14 Among them, Porphyromonas gingivalis (Pg) is an established keystone pathogen capable of orchestrating community-wide dysbiosis even at low abundance.^15,16 P. gingivalis promotes tumor progression through multiple mechanisms, including modulation of apoptosis, cell cycle regulation, and tumor microenvironmental inflammation.^17,18 Notably, P. gingivalis infection can also suppress antitumor immunity by promoting regulatory T cell differentiation and increased immune checkpoint activity such as PD-L1 expression on dendritic cells.^10,19 These multifaceted effects position P. gingivalis as a potential driver of malignancies across the orodigestive tract, including esophageal squamous cell carcinoma (ESCC).

ESCC is a highly aggressive malignancy with marked geographic variability and poor clinical outcomes, with a five-year survival rate below 20%.^20,21 Despite advances in treatment, early detection remains limited, and most patients are diagnosed at advanced stages. Emerging evidence suggests that the esophageal microbiome may contribute to ESCC pathogenesis; however, its composition and functional role remain incompletely defined. Epidemiological studies have linked specific oral pathogens with esophageal cancer risk, including Tannerella forsythia with esophageal adenocarcinoma (EAC) and P. gingivalis with ESCC.²² Although the anatomical continuity between the oral cavity and esophagus suggests potential linkage,^23,24 prior studies have largely examined these sites independently and have reported inconsistent taxonomic signatures. As a result, it remains unclear whether consistent microbial alterations span both niches, and whether such cross-site patterns have diagnostic or biological relevance in ESCC. Importantly, no prior study has systematically analyzed paired oral and esophageal microbiomes to define consistent cross-site signatures.

To address this gap, we performed paired analysis of oral swabs and esophageal tumor tissues from patients with ESCC to evaluate cross-site microbial consistency, assess diagnostic potential, and investigate the mechanistic contribution of P. gingivalis.

Results

Characteristics of participants

The major demographic and baseline clinical characteristics of patients with ESCC and healthy controls (HCs) are summarized in Table 1. A total of 45 patients with ESCC (26 men and 19 women, aged 45–89 years) and 45 age- and sex-matched HCs were included in the final analysis following rigorous quality control of all samples. All participants were recruited from the same institution within a high-incidence ESCC region. There are no significant differences between ESCC cases and controls regarding age, sex, smoking status, or drinking status, indicating effective matching of study groups and minimizing potential confounding factors. This balanced cohort composition ensures robustness and reliability for downstream microbiome comparisons and statistical analyses.

Table 1.

Baseline characteristics of healthy controls and patients with ESCC

	HC (n = 45)	ESCC (n = 45)	t/χ2	p value
Age (year)	69.27 ± 8.74	70.71 ± 8.89	0.777	0.439
Gender			0.421	0.517
Female	16 (35.6%)	19 (42.2%)
Male	29 (64.4%)	26 (57.8%)
Smoking			1.385	0.239
No	35 (77.8%)	30 (66.7%)
Yes	10 (22.2%)	15 (33.3%)
Drinking			0.809	0.368
No	40 (88.9%)	37 (82.2%)
Yes	5 (11.1%)	8 (17.8%)

Diversity and taxonomic alteration in the oral microbiome of patients with ESCC

High-depth 16S rRNA gene sequencing of oral swabs was performed for all 45 patients with ESCC and 45 matched HCs, with Good’s coverage exceeding 99%, ensuring comprehensive characterization of community diversity. Alpha-diversity analyses show no significant differences in richness or evenness between groups, as measured by Chao1 and Simpson indices (Figures 1A and 1B). In contrast, beta-diversity analyses based on weighted UniFrac PCoA show a statistically significant but modest separation between ESCC and control samples (ANOSIM p = 0.001, R = 0.164), indicating a shift in oral community structure rather than a pronounced clustering pattern (Figure 1C). This pattern is consistent with previously reported oral dysbiosis in carcinogenesis.^19,22

Figure 1.

Diversity, taxonomic composition, and diagnostic profiling of the oral microbiome in patients with ESCC and healthy controls

(A and B) Alpha-diversity analysis using Chao1 and Simpson indices showed no significant differences in richness or evenness between ESCC and control groups.

(C) Beta-diversity based on weighted UniFrac PCoA showed modest but significant separation between ESCC and controls, indicating altered oral microbial community structures.

(D and E) Relative abundance profiling at phylum and genus levels showed enrichment of Firmicutes, Bacteroidota, Fusobacteriota, Streptococcus, Fusobacterium, Alloprevotella, and Porphyromonas in ESCC, with controls showing higher abundance of Proteobacteria, Actinobacteria, Neisseria, Capnocytophaga, Rothia, and Actinomyces.

(F–H) LEfSe analysis identified significantly discriminative taxa at the phylum, genus, and species levels, including increased Streptococcus mitis, Porphyromonas gingivalis, and Alloprevotella tannerae in ESCC samples, contrasting with Streptococcus cristatus and Streptococcus parasanguinis in controls.

(I) Receiver operating characteristic (ROC) analysis for selected genera showed prominent diagnostic value (AUC >0.7 for Capnocytophaga, Leptotrichia, Fusobacterium, Rothia, and others), and the composite panel achieved strong classification performance for ESCC.

Taxonomic profiling identified 27 phyla, 64 classes, 136 orders, 243 families, 550 genera, and 935 species across all samples. At the phylum level, ESCC oral microbiota is dominated by Firmicutes, Bacteroidota, and Fusobacteriota, with lower abundance of Proteobacteria, Patescibacteria, Campylobacterota, and Actinobacteriota compared to controls (Figure 1D). At the genus level, ESCC samples show enrichment of Streptococcus (23.69% vs. 15.92%), Fusobacterium (5.09% vs.1.69%), Alloprevotella (5.33% vs. 1.89%), and Porphyromonas (4.02% vs. 2.21%), whereas controls exhibit higher abundance of Neisseria (7.79% vs. 17.23%), Capnocytophaga (1.64% vs. 4.38%), Rothia (1.14% vs. 3.71%), and Actinomyces (1.26% vs. 1.62%) (Figure 1E; Table 2).

Table 2.

Relative abundances of the top 30 genera of oral microbiota in patients with ESCC and healthy control subjects

Genus	ESCC (%)	HC (%)	p value	q value
Streptococcus	23.69	15.92	0.026	0.099
Neisseria	7.79	17.23	0.010	0.049
Haemophilus	8.83	7.82	0.881	0.930
Veillonella	7.39	9.13	0.926	0.968
Prevotella_7	4.49	3.25	0.292	0.389
Alloprevotella	5.33	1.89	0.003	0.019
Fusobacterium	5.09	1.69	< 0.001	< 0.001
Gemella	3.31	3.06	0.274	0.389
Porphyromonas	4.02	2.21	0.027	0.102
Capnocytophaga	1.64	4.38	< 0.001	< 0.001
Rothia	1.14	3.71	< 0.001	0.003
Prevotella	2.95	1.76	0.071	0.207
Granulicatella	1.11	2.20	0.015	0.065
Leptotrichia	2.52	0.70	< 0.001	< 0.001
Selenomonas	1.34	1.66	0.307	0.389
Aggregatibacter	1.39	1.53	0.210	0.389
Actinomyces	1.26	1.62	0.051	0.164
Treponema	1.14	0.99	0.019	0.080
Campylobacter	0.59	1.44	0.003	0.019
Lautropia	0.59	1.35	0.001	0.006
Actinobacillus	0.80	0.91	0.345	0.420
Centipeda	0.49	0.88	0.115	0.279
Klebsiella	0.13	0.91	0.465	0.550
Trichococcus	0.31	0.70	0.003	0.020
Corynebacterium	0.57	0.40	0.990	0.996
TM7x	0.44	0.46	0.532	0.616
Dialister	0.38	0.48	0.781	0.834
Atopobium	0.15	0.65	0.001	0.010
Abiotrophia	0.26	0.53	< 0.001	0.003
F0332	0.61	0.13	0.460	0.546

Further LEfSe analysis identified 113 taxa with significant differential abundance (LDA > 3, p < 0.05), including 47 enriched in ESCC and 66 in controls (Figures 1F–1H). These results confirm the compositional shifts, reinforcing enrichment of pro-tumorigenic taxa alongside reduction of commensal genera, consistent with the patterns described above. At the species level, Streptococcus mitis, P. gingivalis, and Alloprevotella tannerae are more abundant in ESCC, whereas Streptococcus cristatus and Streptococcus parasanguinis predominated in controls (Figure 1H). Fusobacterium was observed to be enriched in the oral swabbing samples of ESCC, although specific species-level classification was not definitive. Given that multiple species of Fusobacterium and Porphyromonas have been implicated in promoting malignant behavior in tumor cells,^25,26 these findings highlight a pattern of oral dysbiosis in ESCC characterized by the enrichment of potentially pathogenic taxa.

To assess the clinical relevance of oral microbial shifts, we evaluated the discriminative ability of ESCC-associated genera in a diagnostic model (Figure 1I). Fourteen genera show significant associations with ESCC diagnosis and are ranked by their area under the receiver operating characteristic curve (AUC). Pro-tumorigenic (positively associated with ESCC) genera include Leptotrichia (AUC = 0.7674), Fusobacterium (AUC = 0.7635), Alloprevotella (AUC = 0.682), Treponema (AUC = 0.644), Streptococcus (AUC = 0.636), and Porphyromonas (AUC = 0.6351). Anti-tumorigenic (negatively associated with ESCC) genera include Capnocytophaga (AUC = 0.7788), Neisseria (AUC = 0.6578), and Rothia (AUC = 0.7289). Importantly, integrating all these genera into a composite diagnostic model markedly improved classification performance (AUC = 0.9422), suggesting that combined microbial signatures may have stronger discriminative capacity than single taxa within this cohort. These results support the concept of a candidate microbial signature for ESCC. However, given this single-cohort design and lack of independent validation, these findings should be considered preliminary and hypothesis-generating. Future studies will incorporate cross-validation and external validation cohorts to confirm model robustness.

PICRUSt-based functional prediction of oral swab samples reveals significant enrichment of genes associated with purine metabolism, pyrimidine metabolism, amino-sugar and nucleotide-sugar metabolism, and DNA-replication proteins in patients with ESCC (Figure 2A). These predicted functional shifts suggest enrichment of pathways associated with microbial nucleotide metabolism and replication-related function within the dysbiotic oral microbiome of patients with ESCC. Enrichment of peptidase- and glycan degradation-related pathways may indicate altered microbial functional potential associated with mucosal glycan processing. In contrast, pathways linked to glyoxylate and dicarboxylate metabolism and other central-carbon metabolic routes are markedly reduced, suggesting altered representation of pathways potentially associated with microbial adaptation to nutrient stress. Although PICRUSt pathway names include some broad KEGG annotations, the curated interpretation here focuses on bacterial metabolic functions. Collectively, these predicted functional shifts suggest altered microbiome-associated metabolic potential in ESCC-associated oral dysbiosis, potentially contributing to inflammatory and oncogenic processes (Figure 2A). Notably, these findings are based on PICRUSt predictions from 16S rRNA data and should therefore be interpreted as exploratory rather than definitive. Taken together, these findings indicate that the oral microbiome of patients with ESCC exhibits altered community structure, enrichment of pro-tumorigenic taxa, and predicted metabolic shifts, which may collectively contribute to disease progression.

Figure 2.

Curated microbial metabolic pathway alterations predicted by PICRUSt in ESCC oral and tumor microbiomes

(A) PICRUSt-based functional prediction of oral swab microbiomes (n = 90) showed enrichment of pathways related to nucleotide synthesis, cell division, glycan degradation, and peptidase activity in patients with ESCC, alongside reduced glyoxylate/dicarboxylate metabolism compared to healthy controls.

(B) In esophageal tumor tissue samples (n = 37 ESCC, n = 8 healthy controls), notable enrichment was observed for bacterial ion-coupled transporters, amino acid metabolism, N-glycan biosynthesis, and glycosyltransferases, with decreased fatty acid, propanoate, pyruvate, ketone body, and retinol metabolism. Pathways exclusively mapped to host biology by KEGG, or subject to ambiguous assignment, were interpreted with caution and prioritized for microbial relevance. These curated findings suggest the ESCC-associated microbiome contributes to metabolic remodeling, immune modulation, and promotion of tumor progression. Statistical significance was tested by the Wilcoxon rank-sum test.

Diversity and taxonomic alteration in the esophageal microbiome of patients with ESCC

We next asked whether similar taxonomic shifts are present in the esophageal tumor microenvironment. Five-region (5R) 16S rRNA gene sequencing was performed on 37 tumor tissues from patients with ESCC and 8 matched esophageal mucosal biopsies from HCs. Sequencing identified 1,481 microbial species with coverage exceeding >99%. Tumor tissues exhibit significantly greater richness and evenness than normal esophageal mucosa (Chao1 index 152.94 vs. 103.07, p = 0.003; Simpson index 0.91 vs. 0.84, p = 0.009, Figures 3A and 3B). This increased diversity may reflect a more permissive and inflammatory tumor microenvironment capable of supporting broader microbial colonization. Further PCoA analyses based on Bray-Curtis distances show partial overlap between tumor and normal samples, contrasting with the more pronounced clustering seen in oral communities (Bray-Curtis PCoA, ANOSIM p = 0.018, R = 0.194, Figure 3C). These findings suggest that tumorigenesis in the esophagus is associated with subtler but significant community shifts, likely driven by non-indigenous keystone bacteria.

Figure 3.

Diversity, taxonomic composition, and diagnostic potential of the esophageal tumor microbiome in patients with ESCC and healthy controls

(A and B) Alpha-diversity using Chao1 and Simpson indices showed significantly increased richness and evenness in ESCC tumor tissues versus healthy esophageal mucosa.

(C) Beta-diversity analysis (Bray-Curtis PCoA) showed partial separation between tumor and control microbial communities.

(D–F) Phylum-, genus-, and species-level taxonomic profiling indicates enrichment of Bacteroidota, Fusobacteria, Prevotella, Fusobacterium, Porphyromonas, Treponema, and Capnocytophaga in ESCC tumors, with reduced Firmicutes, Streptococcus, Rothia, Neisseria, and Actinomyces.

(G–I) LEfSe analysis identified significantly discriminative genera and species associated with ESCC tumor tissues (e.g., Fusobacterium nucleatum, Capnocytophaga sputigena, Porphyromonas gingivalis, Streptococcus anginosus) and controls (e.g., Streptococcus mitis, Streptococcus pneumoniae, Rothia mucilaginosa, Actinomyces odontolyticus).

(J and K) ROC analyses of key genera and species showed robust diagnostic value for ESCC (AUC > 0.7 for Fusobacterium, Prevotella, Porphyromonas, Capnocytophaga, and Streptococcus, among others), with integration into composite panels yielding high diagnostic accuracy. “∗∗” indicates p < 0.01.

At the phylum level, Bacteroidota, Fusobacteria, and Spirochaetes are enriched in tumors, whereas Firmicutes, Proteobacteria, and Actinobacteria are more abundant in normal mucosa (Figure 3D). These patterns mirror the increased presence of anaerobic, inflammation-associated phyla that often dominate hypoxic tumor environments. At the genus level, tumors show enrichment of Prevotella, Veillonella, Fusobacterium, Capnocytophaga, Porphyromonas, and Treponema, with reduced Streptococcus, Neisseria, Rothia, and Actinomyces relative to controls (Figure 3E; Table 3). Consistently, species-level analysis further highlights the enrichment of Fusobacterium nucleatum, Prevotella melaninogenica, Capnocytophaga sputigena, and P. gingivalis in tumor tissues, whereas Streptococcus mitis, Neisseria subflava, and Rothia mucilaginosa are more abundant in normal mucosa (Figures 3F and 3I; Table 4). These hierarchical shifts from phylum to species level underscore a transition toward anaerobic, proteolytic, and potentially oncogenic bacteria within the esophageal tumor niche. LEfSe analysis further confirmed these compositional shifts and identified tumor-associated taxa consistent with those observed in oral samples (Figures 3G–3I).

Table 3.

Relative abundances of the top 30 genera of esophageal microbiota in patients with ESCC and healthy control subjects

Genus	ESCC (%)	HC (%)	p value	q value
Streptococcus	13.10	42.07	< 0.001	0.011
Prevotella	20.17	6.18	0.002	0.039
Neisseria	6.42	12.50	0.022	0.183
Actinobacillus	3.99	7.17	0.835	0.906
Veillonella	8.19	2.89	0.357	0.748
Fusobacterium	7.42	1.99	0.019	0.177
Pseudomonas	0.15	5.93	0.007	0.080
Haemophilus	4.46	1.52	0.406	0.834
Capnocytophaga	4.21	0.07	0.003	0.039
Selenomonas	3.29	0.12	0.003	0.039
Achromobacter	0.00	3.23	< 0.001	< 0.001
Enterococcus	2.80	0.00	0.188	0.597
Rothia	0.14	2.09	0.001	0.022
Gemella	0.78	1.21	0.458	0.840
Catonella	1.95	0.03	0.001	0.022
Porphyromonas	1.57	0.38	0.012	0.123
Treponema	1.92	0.02	0.012	0.123
Aggregatibacter	1.77	0.17	0.107	0.412
Leptotrichia	1.29	0.39	0.187	0.603
Providencia	1.55	0.00	0.228	0.656
Granulicatella	0.59	0.96	0.079	0.343
Helicobacter	0.09	1.39	0.807	0.885
Peptostreptococcus	1.12	0.09	< 0.001	0.005
Campylobacter	0.93	0.27	0.523	0.786
Actinomyces	0.21	0.91	0.016	0.154
Bulleidia	0.33	0.75	0.485	0.879
Dialister	0.90	0.11	0.117	0.445
Lachnoanaerobaculum	0.71	0.09	0.003	0.039
Escherichia_Shigella	0.66	0.00	0.021	0.180
Oribacterium	0.45	0.16	0.161	0.533

Table 4.

Relative abundance of bacterial species in esophageal tissues from patients with ESCC and healthy control subjects

Species	HC (%)	ESCC (%)	p value	q value
Streptococcus mitis	23.14	6.72	0.001	0.099
Streptococcus pneumoniae	10.95	1.80	0.008	0.282
Fusobacterium nucleatum	1.21	3.14	0.047	0.363
Neisseria meningitidis	2.60	0.80	0.042	0.338
Achromobacter piechaudii	3.06	0.00	< 0.001	< 0.001
Rothia mucilaginosa	1.83	0.12	0.001	0.099
Capnocytophaga sputigena	0.00	1.80	0.009	0.310
Peptostreptococcus stomatis	0.09	1.12	< 0.001	0.021
Prevotella oris	0.09	1.07	0.039	0.316
Streptococcus anginosus	0.01	1.05	0.005	0.206
Actinomyces odontolyticus	0.61	0.16	0.014	0.415
Catonella morbi	0.01	0.68	0.005	0.206
Capnocytophaga leadbetteri	0.01	0.59	0.035	0.288
Streptococcus salivarius	0.46	0.11	0.003	0.127
Neisseria cinerea	0.01	0.54	0.044	0.341
Streptococcus cristatus	0.39	0.06	0.034	0.284
Selenomonas sputigena	0.00	0.40	0.012	0.370
Streptococcus peroris	0.27	0.04	< 0.001	0.026
Rothia dentocariosa	0.19	0.01	0.007	0.253
Porphyromonas gingivalis	0.01	0.17	0.036	0.295
Pelomonas puraquae	0.18	0.00	0.032	0.777
Pelomonas aquatica	0.15	0.00	< 0.001	0.012
Achromobacter spanius	0.12	0.00	0.032	0.777
Dechloromonas agitata	0.09	0.00	< 0.001	0.001
Pseudomonas azotoformans	0.09	0.00	0.001	0.099
Bacillus anthracis	0.08	0.00	0.002	0.165
Azospira oryzae	0.08	0.00	0.032	0.777
Acinetobacter guillouiae	0.08	0.00	0.002	0.165
Rothia aeria	0.07	0.01	0.010	0.322
Eikenella corrodens	0.00	0.07	0.009	0.310
Roseateles depolymerans	0.06	0.00	0.032	0.777
Pseudomonas putida	0.05	0.01	< 0.001	0.021
Massilia yuzhufengensis	0.06	0.00	0.032	0.777
Azospirillum brasilense	0.06	0.00	0.002	0.165
Blastomonas natatoria	0.06	0.00	< 0.001	< 0.001
Massilia timonae	0.05	0.00	0.002	0.165
Achromobacter denitrificans	0.05	0.00	0.002	0.165
Mycobacterium phocaicum	0.05	0.00	0.032	0.777
Aerococcus urinaeequi	0.04	0.00	0.032	0.777
Actinomyces meyeri	0.04	0.00	0.032	0.777
Bacillus thuringiensis	0.04	0.00	0.032	0.777
Microbacterium chocolatum	0.04	0.00	0.002	0.165
Cupriavidus campinensis	0.04	0.00	0.002	0.165
Pseudomonas libanensis	0.04	0.00	0.032	0.777
Methyloversatilis universalis	0.03	0.00	0.032	0.777
Rhizobium radiobacter	0.03	0.00	0.006	0.228
Alcaligenes faecalis	0.03	0.00	0.032	0.777
Hydrotalea flava	0.03	0.00	0.032	0.777
Comamonas testosteroni	0.03	0.00	0.002	0.165
Azotobacter chroococcum	0.03	0.00	0.032	0.777
Acinetobacter lwoffii	0.02	0.00	0.032	0.777
Novosphingobium stygium	0.02	0.00	0.032	0.777
Lactococcus piscium	0.02	0.00	0.032	0.777
Pseudomonas plecoglossicida	0.02	0.00	0.032	0.777
Acinetobacter parvus	0.02	0.00	0.032	0.777
Pseudomonas protegens	0.02	0.00	0.002	0.165
Ralstonia mannitolilytica	0.02	0.00	0.032	0.777
Comamonas terrigena	0.02	0.00	0.032	0.777
Burkholderia contaminans	0.01	0.00	0.032	0.777
Lysinibacillus boronitolerans	0.01	0.00	0.032	0.777
Acinetobacter baumannii	0.01	0.00	0.032	0.777
Bacillus weihenstephanensis	0.01	0.00	0.032	0.777
Variovorax paradoxus	0.01	0.00	0.032	0.777
Sphingomonas roseiflava	0.01	0.00	0.032	0.777
Clostridium celerecrescens	0.01	0.00	0.032	0.777
Diaphorobacter nitroreducens	0.01	0.00	0.002	0.165
Actinomyces dentalis	0.01	0.00	0.020	0.165
Mycobacterium mucogenicum	0.01	0.00	0.032	0.777
Geobacillus vulcani	0.01	0.00	0.032	0.777
Staphylococcus cohnii	0.01	0.00	0.032	0.777
Methylobacterium zatmanii	0.01	0.00	0.032	0.777
Pseudomonas brenneri	0.01	0.00	0.032	0.777
Paracoccus marcusii	0.01	0.00	0.032	0.777

Diagnostic modeling again showed that bacterial genera could discriminate ESCC from controls. Six specific genera show consistent diagnostic value in both oral and esophageal sites, including increased Fusobacterium (AUC = 0.7669), Prevotella (AUC = 0.848), Porphyromonas (AUC = 0.7872), and Treponema (AUC = 0.777), and decreased Rothia (AUC = 0.8666) and Neisseria (AUC = 0.7601) in ESCC tumor tissues (Figures 3J and 3K). Similar to the oral swabbing samples, a combined panel yielded strong diagnostic power (AUC > 0.90) (Figures 3J and 3K), although this should be interpreted cautiously given the limited sample size and absence of independent validation. Interestingly, Capnocytophaga and Streptococcus show opposite trends between oral and esophageal samples, highlighting the impact of microenvironment and species specificity on microbiome structure.

PICRUSt-based functional prediction infers relative enrichment of metabolism-related pathways in tumor tissues, notably for ion-coupled transporters, N-glycan biosynthesis, select amino acid metabolic pathways, glycosyltransferases, and epithelial cell signaling in Helicobacter pylori infection (Figure 2B). Some KEGG categories, such as prostate cancer, progesterone-mediated oocyte maturation, and antigen processing, likely reflect annotation overlap rather than direct bacterial function and are excluded from the primary interpretation of microbiome-associated metabolic pathways.

We also observed reduced representation of pathways related to fatty acid metabolism, pyruvate metabolism, branched-chain amino acid biosynthesis, ketone body metabolism, propanoate and lysine metabolism, and retinol metabolism. Although KEGG may map microbial genes to broad metabolic categories, our interpretation focused on pathways with clear bacterial origin or relevance. Notably, reduced abundance of predicted fatty acid- and propanoate-related pathways suggests altered microbial functional potential associated with lipid mediators and short-chain fatty acid metabolism, which are important for epithelial integrity and immune homeostasis within the tumor microenvironment.^27,28 These predicted pathway alterations may therefore be associated with inflammatory and immune changes relevant to ESCC progression. Enrichment of microbial N-glycan biosynthesis and glycosyltransferase-related pathways reflects altered microbial functional potential associated with glycosylation, with possible implications for immune evasion and tumor growth.²⁹ Likewise, altered representation of ion-coupled transporter-related pathways may be associated with changes in the tumor ionic and metabolic environment, with potential implications for extracellular pH and redox balance.³⁰ Collectively, these predicted functional shifts suggest altered microbiome-associated metabolic potential in ESCC tumors, likely modifying inflammatory and immune processes relevant to disease progression. However, as these results are based on PICRUSt inference from 16S rRNA data, they should be interpreted as predictive and exploratory and will require further validation using direct functional approaches.

Associations between oral and esophageal microbiota in ESCC

To investigate cross-site microbiome alterations, we compared paired oral and esophageal tumor samples from the same patients with ESCC. Quality-controlled datasets yield 1,340 and 1,300 microbial species in oral and esophageal samples, respectively, with 823 shared between the two niches (Figure 4A), suggesting substantial cross-compartmental microbial overlap. Alpha-diversity analyses show higher community richness (Chao1: 260.87 vs. 153.66; p < 0.0001; Figure 4B) and evenness (Simpson index; Simpson: 0.93 vs. 0.91; p < 0.001; Figure 4C) in oral versus esophageal tumor microbiota, suggesting that the oral environment maintains a more diverse and balanced community, while esophageal tumors are more selective and susceptible. This difference may underlie the more significant impact of microbiome shifts observed in ESCC compared to oral cancer. Further PCoA analyses show tighter clustering of oral communities relative to esophageal ones (Bray-Curtis PCoA, ANOSIM, p = 0.001), reflecting lower inter-individual variability and possible colonization by oral indigenous microbes translocated to the esophageal tumor niche (Figure 4D). The substantial oral-esophageal overlap supports the hypothesis that the oral cavity can act as a reservoir for tumor-associated bacteria.

Figure 4.

Cross-site comparative analysis of paired oral and esophageal tumor microbiota in patients with ESCC

(A) Venn diagram shows the number of shared and unique microbial species between oral swab (ESCC-S) and esophageal tumor (ESCC-T) samples, highlighting substantial overlap across compartments.

(B and C) Alpha-diversity analyses show significantly greater richness (Chao1 index) and evenness (Simpson index) in the oral microbiome compared to paired esophageal tumor microbiota.

(D) Beta-diversity analysis (Bray-Curtis PCoA) shows tighter clustering of oral communities and greater heterogeneity in esophageal tumor microbiota.

(E–G) Relative abundance barplots at phylum, genus, and species levels showed shared and niche-specific taxa across oral and tumor sites.

(H) Cross-site correlation matrix shows significant positive associations for species including Granulicatella elegans, Fusobacterium periodonticum, Porphyromonas gingivalis, and Prevotella tannerae, indicating coordinated abundance of key pro-tumorigenic taxa.

(I) Intersection analysis identified Porphyromonas gingivalis as the only species shared across oral-enriched, esophageal-enriched, and cross-correlated taxa in ESCC. “∗∗∗” and “∗∗∗∗” indicate p < 0.001 and p < 0.0001, respectively.

At the phylum level, Firmicutes, Bacteroidota, and Proteobacteria predominate in both niches, with differences in Fusobacteriota, Actinobacteriota, and TM7 (Figure 4E). At the genus level, Streptococcus, Prevotella, Veillonella, and Neisseria are prevalent in both sites, although Prevotella predominates in the esophagus, and Streptococcus predominates in the oral cavity (Figure 4F). Cross-site correlation analyses identified multiple positively associated (ρ > 0.4, p < 0.05) species, including Granulicatella elegans, Fusobacterium periodonticum, P. gingivalis, and Prevotella tannerae (Figure 4H; Table 5), confirming coordinated abundance patterns across niches and implicating pro-tumorigenic taxa (e.g., F. nucleatum, P. gingivalis) in cross-site ecology. These patterns further suggest that the oral cavity may serve as a reservoir for tumor-associated bacteria. Collectively, these findings highlight the potential of non-invasive oral sampling to reflect tumor-associated microbial signals and aid in risk stratification/monitoring.

Table 5.

Correlation analysis of the top 30 species between the oral microbiota and the esophageal microbiota in patients with ESCC

Species	ESCC_S	ESCC_T	ρ	p value
Granulicatella elegans	0.73	0.14	0.759	< 0.001
Veillonella dispar	4.961	5.28	0.525	0.001
Neisseria macacae	5.33	1.47	0.556	0.001
Prevotella intermedia	1.88	4.82	0.513	0.002
Haemophilus haemolyticus	2.06	1.24	0.512	0.002
Neisseria meningitidis	0.66	0.85	0.51	0.002
Porphyromonas gingivalis	1.05	0.18	0.505	0.002
Treponema amylovorum	0.27	0.78	0.496	0.002
Fusobacterium periodonticum	0.61	3.01	0.493	0.003
Haemophilus segnis	0.93	0.65	0.472	0.004
Prevotella tannerae	2.10	1.24	0.435	0.009
Neisseria mucosa	0.87	0.29	0.405	0.016
Neisseria subflava	1.84	2.49	0.392	0.020
Peptostreptococcus stomatis	0.23	1.17	0.349	0.040
Capnocytophaga leadbetteri	0.76	0.61	0.333	0.050
Streptococcus mitis	10.93	7.09	0.289	0.092
Streptococcus pneumoniae	2.20	1.83	0.266	0.122
Streptococcus oralis	0.84	0.12	0.246	0.155
Veillonella parvula	2.27	2.65	0.201	0.247
Haemophilus parainfluenzae	5.28	2.21	0.194	0.263
Prevotella pallens	0.45	1.22	0.191	0.273
Aggregatibacter segnis	1.01	0.46	0.187	0.282
Granulicatella adiacens	0.72	0.40	0.184	0.291
Capnocytophaga sputigena	0.50	1.89	0.104	0.554
Prevotella nigrescens	0.71	0.46	0.101	0.562
Prevotella oris	0.21	1.13	0.078	0.655
Prevotella melaninogenica	4.11	5.99	−0.048	0.783
Neisseria perflava	0.55	0.46	0.015	0.930
Streptococcus anginosus	0.07	0.94	0.009	0.960
Fusobacterium nucleatum	1.75	3.10	0.003	0.987

Additionally, we examined associations between tumor-enriched species and clinicopathologic features. C. sputigena is associated with metastasis, and several species including S. pneumoniae, A. odontolyticus, and P. gingivalis are significantly more abundant in advanced-stage ESCC (Figures 5A and 5B). An intersection analysis integrating oral-enriched taxa, esophageal-enriched taxa, and cross-site correlated taxa identified P. gingivalis as the sole species shared across all three sets of patients with ESCC (Figure 4I). This repeated detection across niches, together with findings from previous studies,^5,19,31 supports the view that P. gingivalis is a keystone bacterium in ESCC-associated dysbiosis and linked to ESCC progression, prompting us to further investigate its functional mechanisms in an animal model.

Figure 5.

Associations between key bacterial species and clinicopathologic features in ESCC tumor tissues

(A) Heatmap shows the relative abundance of 20 representative bacterial species across clinical variables including gender, age, smoking, drinking, nodal status, tumor (T), node (N), and TNM stage.

(B–E) Boxplots show significant differences in relative abundance for specific species between patient subgroups. (B) Capnocytophaga sputigena by nodal metastasis (N0 vs. N1-2), (C) Actinomyces odontolyticus by tumor stage (T2 vs. T3), (D) Streptococcus pneumoniae by tumor stage (T2 vs. T3), and (E) Porphyromonas gingivalis by early (I + II) versus advanced (III) TNM stage. “∗” indicate statistically significant differences (∗p < 0.05).

Porphyromonas gingivalis promotes esophageal squamous cell carcinogenesis in a 4-NQO-induced mouse model

Consistent identification of P. gingivalis in human ESCC samples suggests a possible role in disease-associated dysbiosis. We therefore investigated its effects in a 4-NQO-induced mouse model of ESCC (Figure 6A). P. gingivalis alone does not alter esophageal appearance or histology. However, combined exposure with 4-NQO results in dramatically thickened, roughened esophageal walls and increased papillomatous lesions compared to 4-NQO alone (Figures 6B and 6C). Histopathology confirmed widespread severe dysplasia, epithelial thickening, and deeper muscle layer infiltration in the combined group. ESCC incidence reaches 100% (6/6) in the P. gingivalis positive/4-NQO group, showing a trend toward increased incidence compared to untreated (0/11), P. gingivalis only (0/5), and 4-NQO alone (54.5%, 6/11) (Table S1) (Fisher’s exact test, p = 0.075). Lesion area is also larger in the colonized group (1.32 mm² vs. 0.25 mm², p < 0.01, Figure 6C). Given prior evidence linking P. gingivalis to PD-L1 expression on dendritic cells and STAT3 activation, we examined the STAT3-PD-L1 axis as a candidate immune pathway. Immunohistochemistry shows increased p-Stat3 staining in the combined group (Figures 6D and 6E), which is further supported by western blot analysis showing significantly higher p-STAT3 expression in the P. gingivalis plus 4-NQO group, as compared to other groups (Figures 6F and 6G). Cytokine analyses also show increased expression of IFN-γ, TNF-α, and IL-6, pointing to a heightened tumor-promoting inflammatory state (Figures 6H–6J). Additionally, flow cytometry analyses demonstrate significantly higher PD-L1 expression on dendritic cells in cervical lymph nodes in the P. gingivalis-positive/4-NQO group compared to the control groups, both in the percentage of positive cells (CD11c⁺/PD-L1⁺) and MFI (Figures 6K–6M), indicating that P. gingivalis may suppress cytotoxicity of CD8⁺ T cells and promote immune escape in ESCC. Collectively, these findings support a tumor-promoting role for P. gingivalis in ESCC, potentially involving microbiome-associated functional alterations as well as the modulation of inflammatory and tumor immune responses.

Figure 6.

Porphyromonas gingivalis accelerates 4-NQO-induced esophageal tumorigenesis by promoting inflammation and immune suppression

(A) Schematic illustration of the experimental design. C57BL/6 mice (n = 11) were administered 4-NQO (50 μg/ml) in drinking water for 8 weeks to induce esophageal carcinogenesis and orally inoculated with P. gingivalis (10¹⁰ CFU/mL) or vehicle control as indicated. Only mice with confirmed oral P. gingivalis colonization were included in subsequent analyses (final n = 5 for Pg only and n = 6 for 4-NQO + Pg groups).

(B) Representative gross morphology and H&E staining of esophageal tissues from control, P. gingivalis (Pg) alone, 4-NQO alone, and 4-NQO + Pg groups after 34 weeks. Combined treatment resulted in extensive epithelial hyperplasia, dysplasia, and invasive carcinoma. Scale bars, 0.2 mm (low magnification) and 0.05 mm (high magnification).

(C) Quantification of total precancerous lesion area per esophagus showed significantly increased lesion burden in the 4-NQO + Pg group compared with 4-NQO alone.

(D and E) Immunohistochemical staining of phosphorylated STAT3 (p-STAT3) in esophageal tissues showed stronger nuclear staining and higher IHC scores in the 4-NQO + Pg group, consistent with activation of STAT3 signaling.

(F and G) Western blot analysis of phosphorylated STAT3 (p-STAT3) expression in esophageal tissues and its quantification normalized to GAPDH.

(H–J) qRT-PCR analysis of IFN-γ, TNF-α, and IL-6 mRNA levels in esophageal tissues showed marked upregulation of pro-inflammatory cytokines following P. gingivalis colonization under 4-NQO exposure.

(K–M) Flow cytometry analysis of draining lymph-node dendritic cells (CD11c⁺) showed increased PD-L1 expression and higher MHC II mean fluorescence intensity (MFI) in the 4-NQO + Pg group, indicating enhanced immune-suppressive potential. “∗”, “∗∗”, and “∗∗∗” represent p < 0.05, p < 0.01, and p < 0.001, respectively.

Discussion

Although oncogenic bacteria and overall microbiome balance in oral and esophageal cancers are increasingly recognized, studies have reported conflicting microbial signatures and cross-site colonization. Previous work has usually focused on a single anatomical site or used different methodological approaches,^32,33 leaving the consistency and the prognostic and/or diagnostic value of cross-site microbial alterations unclear. In this study, we utilized paired oral and esophageal tumor samples from patients with ESCC and HCs, identifying a consistent set of signature bacterial genera including increased Porphyromonas, Fusobacterium, and Treponema, and decreased Neisseria, Rothia, and Actinomyces across both sites. Importantly, these cross-site microbial signatures are strongly associated with ESCC status in this cohort, and that P. gingivalis may serve as a key oral bacterium promoting tumor development by driving dysbiosis and creating a tumor-permissive microenvironment. Together, these findings suggest that consistent microbiome alterations, centered on key pathogenic taxa, bridge the oral cavity and the esophageal tumor niche, providing mechanistic and translational insights into ESCC pathogenesis and disease stratification.

A notable finding is the consistently higher microbial diversity in the oral cavity than in the esophageal mucosa in both patients with ESCC and controls. Specifically, oral samples show significantly greater richness and evenness (alpha diversity) (Figure 4B), along with stronger community separation (beta diversity) (Figure 4C), reflecting a more complex and resilient microbial ecosystem. Higher oral microbiome diversity is generally linked to ecological stability and resistance to perturbations, which may help buffer pathobiont overgrowth and maintain homeostasis.^34,35 This stability may partly explain the inconsistent identification of specific oncogenic bacteria in oral cancer studies, where a diverse and competitive microbial community may limit dominance by any single pathogenic species. In contrast, the reduced diversity identified in the esophageal microbiome, especially within tumor tissues, likely creates an ecological niche more vulnerable to colonization by introduced pathogens such as P. gingivalis and F. nucleatum.^19,36 Lower richness and evenness in the esophagus may therefore facilitate pathogen persistence and contribute to carcinogenesis by amplifying inflammatory and microbiome-associated functional disturbances, consistent with findings from previous studies.^37,38 These findings highlight the importance of diversity-driven ecological dynamics across anatomical sites in ESCC and warrant further study of how loss of microbial diversity predisposes to pathogen-mediated tumorigenesis and affects therapeutic response.

Although we identified Porphyromonas, Fusobacterium, and Treponema as key pro-tumorigenic genera in ESCC, other studies have reported different combinations, including Porphyromonas, Prevotella, Treponema, or Fusobacterium, Streptococcus, Veillonella.^39,40 Certain taxa including Streptococcus can be oncogenic in certain contexts but protective in others,^41,42 underscoring that bacterial functions are highly context-dependent and influenced by host-microbe interactions. More broadly, “health” is not defined by a single fixed community composition but by a dynamic spectrum of microbial states influenced in part by host immunity. The immune system imposes selective pressures on colonizing bacteria, thereby shaping microbial patterns most compatible with mucosal homeostasis.^41,43 In addition to host factors, methodological variation also contributes to discrepancies among studies, including differences in sampling site and depth, disease stage, and DNA/RNA extraction and sequencing strategies. To minimize these confounders, we used a paired case-control design and high-resolution 5R 16S rRNA sequencing to esophageal cancer tissues, which improved taxonomic resolution and reduced classification ambiguity.⁴⁴ This approach enables robust identification of Porphyromonas, Fusobacterium, and Treponema as consistent, cross-site ESCC-associated genera with strong diagnostic performance. These findings are consistent with their reported roles in other GI and oral cancers^6,10,45 and support further validation in larger, independent cohorts.

In addition to tissue-associated microbiota, emerging evidence also implicates gut microbiome alterations in ESCC progression and therapeutic response.^46,47 However, tissue-associated microbiota and fecal/gut microbiota differ substantially in sampling source, biological interpretation, and clinical relevance. Tissue-associated microbiota obtained from oral swabs or tumor biopsies more directly reflect local mucosal colonization, tumor-associated biofilms, and host-microbe interactions within the tumor immune microenvironment. In contrast, fecal microbiota primarily represents luminal intestinal microbial communities and may better capture systemic metabolic and immune influences on cancer development and treatment response. Therefore, these two microbial compartments likely provide complementary rather than interchangeable biological information in ESCC. Future studies integrating tissue-associated and gut microbiome profiling may help clarify how local microbial ecology and systemic microbiome-associated effects jointly contribute to ESCC pathogenesis and therapeutic outcomes.

P. gingivalis is increasingly recognized not only as an oral keystone pathogen that directly promotes tumorigenesis, but also as a master regulator that shapes the tumor microenvironment through effects on microbial community dynamics and host immune responses.^5,16,48,49 It promotes recruitment and expansion of myeloid-derived suppressor cells (MDSCs) while suppressing cytotoxic T cell activity,^10,50 thereby creating an immunosuppressive milieu that facilitates microbial persistence and tumor immune evasion. It also induces proinflammatory cytokines such as IL-6 and IL-8, along with chemokines that recruit additional tumor-promoting immune cells and activate oncogenic pathways including STAT3,^19,51,52 which are important for tumor progression. Beyond immune modulation, P. gingivalis has been shown to promote epithelial-mesenchymal transition (EMT), inhibit ferroptosis, and enhance invasiveness in oral and other squamous cell carcinomas.⁵³ It should be noted that the murine oral microbiome differs from that of humans, and P. gingivalis is not indigenous to mice. Therefore, its effects in this model may reflect both direct host interactions and indirect modulation of microbial community structure. As such, the 4-NQO model provides mechanistic insight but does not fully recapitulate the human orodigestive microbiome. Collectively, our findings support a role for P. gingivalis in accelerating esophageal carcinogenesis through direct tumor-promoting activity as well as indirect modulation of the immune and microbial milieu.

Recent evidence suggests that pro-tumorigenic bacteria can reshape tumor metabolism in addition to modulating immune and signaling pathways, thereby promoting carcinogenesis.^54,55 Consistent with this, our functional predictions suggest that ESCC-associated microbiomes may exhibit altered representation of pathways related to glycan-, amino acid-, and fatty acid metabolism, as well as epithelial signaling (Figure 2B). These patterns may help explain how microbial communities could influence local metabolic networks that parallel and potentiate cancer cell metabolism. Such predicted functional shifts may be associated with altered epithelial barrier function, sustained chronic inflammation, and favor immune evasion, thereby reinforcing a tumor-promoting niche.^27,28,30 Enrichment of glycosylation-related pathways further suggests microbiome-associated functional involvement in aberrant signaling and tumor-immune crosstalk.²⁹ Together, these findings suggest that the ESCC microbiome may contribute to carcinogenesis not only through immune modulation but also through microbiome-associated functional shifts potentially linked to sustained dysbiosis, inflammation, and therapeutic resistance.

P. gingivalis may also contribute to microbiome-associated functional and ecological alterations potentially affecting local metabolite composition. These functional alterations could influence nutrient conditions that facilitate the expansion of other pro-tumorigenic taxa, such as F. nucleatum and F. periodonticum, which may cooperate with carcinogenic exposures to promote EMT and tumor progression.^56,57 Moreover, P. gingivalis produces gingipains that further degrade host defense molecules and generate metabolites that enhance tumor cell stemness and chemoresistance.^19,58,59 In contrast, commensal bacteria with anti-inflammatory and barrier-protective functions may counteract tumor growth by detoxifying carcinogens, suppressing pro-tumorigenic metabolites, and preserving mucosal metabolic homeostasis.^54,60 Disruption of this beneficial network by P. gingivalis removes key metabolic and ecological safeguards, tipping the balance toward carcinogenesis. Therefore, these effects likely reflect both direct pathogen activity and broader microbiome-mediated interactions. Collectively, our findings support a model in which P. gingivalis contributes to ESCC progression by reshaping microbial ecology, host inflammation, and potentially tumor-associated metabolic pathways. It remains an important priority to further clarify the molecular mechanisms linking pathogenic microbiome-associated functional shifts, including the enrichment of glycan degradation-related pathways and depletion of lipid- and short-chain fatty acid-associated pathways, with ESCC progression.

One intriguing observation is the opposing abundance patterns of Capnocytophaga between oral swabs and esophageal tumor tissues. In the oral microbiome, genus-level analysis reveals Capnocytophaga as negatively associated with ESCC, suggesting a potential protective or anti-tumorigenic role. In contrast, high-resolution 16S rRNA sequencing of tumor tissues shows C. sputigena as significantly enriched in tumors, consistent with evidence that specific Capnocytophaga species can promote EMT and tumor progression within the carcinoma niche.⁶¹ This discrepancy likely reflects the limited resolution of genus-level oral profiling, which pools both protective and pathogenic species, versus species-level tumor analysis, where C. sputigena predominates. Differences in local microenvironments, including immune activity, nutrient availability, and host-microbe interactions, may further shape the functional outcomes of Capnocytophaga colonization. A similar pattern was also observed for Streptococcus, which includes both tumor-suppressive and tumor-promoting species, underscoring the need for species-level, context-aware microbiome analyses. Overall, these findings highlight the ecological and functional complexity of microbiome-cancer relationships and the importance of integrating high-resolution taxonomic and functional profiling to resolve genus-species discrepancies.

In this study, we identified signature bacterial species and microbiome patterns consistently present in both oral swabs and esophageal tumors of patients with ESCC. These cross-site microbial signatures, including pathogens such as P. gingivalis, F. nucleatum, and C. sputigena, deepen understanding of ESCC pathogenesis and show potential prognostic relevance. With further validation in larger cohorts, these taxa and their composite community profiles could serve as non-invasive biomarkers to identify high-risk individuals and predict clinical outcomes. Our recent findings that P. gingivalis infection confers chemoresistance in ESCC further highlight its clinical significance.^6,19 Building on these insights, we propose a microbiome-based “diagnostic code,” in which bacterial signatures and diversity metrics are integrated into a concise microbial barcode processed by machine-learning algorithms or meta-transcriptomic assays. Such an approach could enable rapid, accurate, and non-invasive ESCC screening using oral swab or saliva samples, particularly valuable in high-incidence or resource-limited regions. Similar microbiome-based diagnostic frameworks have been explored for colorectal and cervical cancers,⁶² and advances in high-throughput sequencing, bioinformatics, and artificial intelligence now make scalable clinical implementation feasible. Together, our findings underscore the potential of consistent oral-esophageal microbiome alterations as predictive biomarkers for early ESCC detection and intervention. Future studies incorporating independent external cohorts, longitudinal sampling, and higher-resolution approaches such as shotgun metagenomics or metatranscriptomics will be important to validate and extend these findings.

In summary, this study provides comprehensive evidence that consistent microbiome alterations link the oral cavity and esophageal tumor microenvironment in ESCC. Using paired-site profiling and high-resolution sequencing, we identified a robust set of key pathogens, P. gingivalis, Fusobacterium nucleatum, and Capnocytophaga sputigena, that define the disease-associated microbiome, correlate with prognosis, and support the feasibility of accurate and non-invasive detection. Integrated ecological, predicted functional, and experimental analyses suggest that these pathogens may link the two anatomical sites by reshaping microbial communities, modulating host immunity, and potentially influencing tumor-associated metabolic processes relevant to disease progression. Mechanistic studies further demonstrate that P. gingivalis promotes a tumor-permissive, immunosuppressive niche that supports microbial persistence and malignancy. Collectively, these findings provide a framework for developing a microbiome-based “bacterial diagnostic code,” potentially enhanced by machine learning and meta-transcriptomics, for saliva- or swab-based early detection, risk stratification, and targeted intervention in ESCC.

Limitations of the study

There are some limitations in this study. First, diagnostic modeling was performed in a single cohort without independent external validation, which may limit generalizability and increase the risk of overfitting. Although the observed AUC values are high, they should be interpreted as preliminary. Larger multi-center cohorts with cross-validation and independent testing will be needed to confirm diagnostic utility. Second, this study was conducted in a single high-incidence ESCC region, and geographic, environmental, dietary, and lifestyle factors may influence microbiome composition. Therefore, broader validation across diverse populations is necessary. Third, the current experimental design does not fully distinguish direct pathogen-driven effects of P. gingivalis from broader microbiome-mediated mechanisms, so targeted microbial or gnotobiotic approaches will be required to resolve causality more precisely. Fourth, functional inferences were based on PICRUSt predictions from 16S rRNA data and therefore should be considered exploratory rather than definitive, pending validation by metagenomic or metabolomic approaches. Fifth, more comprehensive immune profiling beyond the STAT3-PD-L1 axis (e.g., cytotoxic CD8⁺ T cell function and MDSC populations), together with targeted modulation of P. gingivalis (e.g., depletion or inhibition strategies) would further clarify causal mechanisms. Sixth, the relatively small number of healthy esophageal biopsy controls may have limited statistical power and generalizability and may also have reduced power to detect additional cross-site correlated taxa. Finally, because biopsies were collected through the oral route, complete exclusion of procedural carryover is not possible despite careful sampling procedures.

Resource availability

Lead contact

All requests for information and resources should be directed to and will be fulfilled by the lead contact, Shegan Gao (gsg112258@163.com).

Materials availability

This study did not generate new unique reagents. Additional information regarding materials and experimental procedures is available from the lead contact upon request.

Data and code availability

• •The 16S rRNA sequencing data reported in this study have been deposited in the Genome Sequence Archive (GSA) under accession nos. CRA036460 (https://ngdc.cncb.ac.cn/gsa/browse/CRA036460) and CRA036462 (https://ngdc.cncb.ac.cn/gsa/browse/CRA036462) are publicly available.
• •This paper does not report original code.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

We acknowledge Dr. Fuyou Zhou (Anyang Tumor Hospital) for providing valuable assistance during the study.

The joint Fund of the Major Science and Technology Projects in Henan Province (grant no. 235200810002); Henan Provincial Key International Science and Technology Cooperation Projects (grant no. 261111520900); The Medical Science and Technology Research Project in Henan Province (grant no. LHGJ 20250560).

Author contributions

J.-Y.K. performed experiments, organized and analyzed data, and prepared figures. Y.-J.Q. conceived the study, provided resources and funding, performed data analysis, and revised the manuscript. Q.-W.L., H.-J.X., D.-R.W., Y.-L.J., P.C., B.-L.G., and Y.-W.L. performed experiments. H.W. and R.J.L. contributed to study conception and manuscript revision. H.W. contributed to study conception and manuscript editing. S.-G.G. conceived and supervised the study, secured funding and resources, organized the research team, and wrote and revised the manuscript. J.-Y.K. and Y.-J.Q. contributed equally to this work. H.W. and S.-G.G. are corresponding authors.

Declaration of interests

The authors declare no competing interests.

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

The authors declare not using any generative AI tools in scientific writing except improve the language of the manuscript.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
CD16/CD32 Monoclonal Antibody (93)	eBioscience (Thermo Fisher Scientific)	Cat#14-0161-82; RRID:AB_467133
CD11c Monoclonal Antibody (N418), PerCP-Cyanine5.5	eBioscience (Thermo Fisher Scientific)	Cat#45-0114-82; RRID:AB_925727
CD274 (PD-L1, B7-H1) Monoclonal Antibody (MIH5), PE	eBioscience (Thermo Fisher Scientific)	Cat#12-5982-82; RRID:AB_466089
Armenian Hamster IgG Isotype Control (eBio299Arm), PerCP-Cyanine5.5	eBioscience (Thermo Fisher Scientific)	Cat#45-4888-80; RRID:AB_906260
Rat IgG2a kappa Isotype Control (eBR2a), PE	eBioscience (Thermo Fisher Scientific)	Cat #12-4321-80; RRID:AB_1834380
Phospho-Stat3 Rabbit Monoclonal Antibody	Cell Signaling Technology	Cat#9145S; RRID:AB_2491009
Bacterial and virus strains
Porphyromonas gingivalis	ATCC	ATCC33277
Biological samples
Human oral swab	This paper	N/A
Human esophageal tissue	This paper	N/A
Chemicals, peptides, and recombinant proteins
4-Nitroquinoline 1-Oxide	Sigma	Cat#8141
TRIzol reagent	Invitrogen	Cat#15596018
AMPure XT beads	Beckman	Cat#A63880
Critical commercial assays
Bacterial Genome DNA Extraction Kit	TIANGEN	Cat#DP302-02
Qubit dsDNA HS Assay Kit	Invitrogen	Cat#Q32854
Universal Two-Step Detection Kit	ZSGB-BIO	Cat#PV9000
DAB Chromogen Kit	ZSGB-BIO	Cat#ZLI-9019
HIScript Ⅱ 1st Strand cDNA Synthesis Kit	Vazyme	Cat#R211-02
AceQ qPCR SYBR Green Master Mix	Vazyme	Cat#Q111-03
Deposited data
16S rRNA gene sequencing data for esophageal tissue	National Genomics Data Center	https://ngdc.cncb.ac.cn/gsa/browse/CRA036460
16S rRNA gene sequencing data for oral swab	National Genomics Data Center	https://ngdc.cncb.ac.cn/gsa/browse/CRA036462
Experimental models: Organisms/strains
Mouse C57BL/6	Cavens	N/A
Oligonucleotides
See Tables S2 and S3 for oligonucleotide list	This paper	Tables S2 and S3
Software and algorithms
FLASH (v1.2.8)	The FLASH team	RRID:SCR_005531; https://ccb.jhu.edu/software/FLASH/
cutadapt (v1.9)	Marcel Martin	RRID:SCR_011841; http://cutadapt.readthedocs.org/en/stable/guide.html
Fqtrim (v0.9.4)	Pertea lab, Johns Hopkins University	RRID:SCR_028291; http://ccb.jhu.edu/software/fqtrim/
Vsearch (v2.3.4)	Rognes lab, University of Oslo	RRID:SCR_024494; https://github.com/torognes/vsearch
QIIME 2 (2019.7)	Caporaso Lab, Northern Arizona University	RRID: SCR_021258; https://qiime2.org/
LEfSe (094f447691f0)	Huttenhower Lab, Harvard T.H. Chan School of Public Health	RRID: SCR_014609; https://huttenhower.sph.harvard.edu/lefse/
PICRUSt2 (v2.3.0)	Langille Lab, Dalhousie University	RRID: SCR_022647; https://github.com/picrust/picrust2
FlowJo (v10)	BD Biosciences	RRID:SCR_008520; https://www.flowjo.com/
MetaboAnalyst 5.0	Wishart Research Group, University of Alberta	RRID:SCR_015539; https://www.metaboanalyst.ca
IBM SPSS Statistics (v26.0)	IBM Corp.	RRID:SCR_016479; https://www.ibm.com/spss
R (v3.4.4)	R Core Team	RRID:SCR_001905; https://www.r-project.org/
Custom R scripts	User-developed	https://www.omicstudio.cn/tool

Experimental models and study participant details

Study participants and ethical approval

Samples were obtained from individuals undergoing routine upper endoscopy from June 2022 to March 2023 at Anyang Cancer Hospital, Henan Province, China, a major referral center located in a high-incidence region for esophageal squamous cell carcinoma (ESCC) that serves patients from multiple regions. Forty-five newly diagnosed (26 men and 19 women, mean age = 70.71 years, SD = 8.89, range: 45–89 years), treatment-naïve ESCC patients and 45 age- and sex-matched healthy controls were enrolled (29 men and 16 women, mean age = 69.27 years, SD = 8.74, range: 48–89 years). Diagnosis was established by histopathological analysis. Among the healthy controls, 8 individuals underwent endoscopy due to clinical symptoms but were subsequently confirmed to have normal esophageal mucosa with no evidence of malignancy. Race, ethnicity, and ancestry data were not collected, as these variables were not considered relevant to the research questions. Inclusion criteria for ESCC cases were: (a) histologically confirmed primary ESCC; (b) no prior anticancer treatment; and (c) no antibiotic or microbiota-modulating medication use within one month before endoscopy. Exclusion criteria included: (a) evidence of metastatic or recurrent malignancy; (b) concurrent malignancy in other organs or severe comorbidities; or (c) recent use of medications affecting the microbiota. For control participants, inclusion criteria additionally required absence of malignancy and no recent antibiotic or microbiota-modulating medication use, with sampling performed under the same pre-endoscopy conditions as ESCC cases (including fasting and restriction of oral hygiene procedures). Demographic and clinical data, including age, sex, smoking history, alcohol consumption, and tumor staging, were retrieved from hospital records. All participants provided written informed consent. The protocol was approved by the Ethics Committee of The First Affiliated Hospital of Henan University of Science and Technology (Approval No. K-2025-B104).

Porphyromonas gingivalis and 4-NQO-induced ESCC mouse model

P. gingivalis strain ATCC 33277 was obtained from ATCC and cultured anaerobically in trypticase soy broth supplemented with yeast extract (1 mg/mL), hemin (5 μg/mL), and menadione (1 μg/mL) as described.¹⁹ Bacteria were grown at 37°C in an AS-500 anaerobic chamber under microaerobic conditions (5% H₂, 5% CO₂, 90% N₂). All animal protocols were approved by the Institutional Animal Care and Use Committee (No. 2022-03-B111). Female C57BL/6 mice (6 weeks old) were randomly assigned to four groups (n = 11 per group): (1) control, (2) P. gingivalis only, (3) 4-NQO only, and (4) 4-NQO + P. gingivalis. 4-NQO (50 μg/mL) was administered in drinking water for 8 weeks, followed by regular water. Mice in the designated groups received oral inoculation with P. gingivalis (1 × 10¹⁰ CFU in PBS) three times weekly from weeks 11–14. Oral swabs were collected at week 20 to confirm successful P. gingivalis colonization by qRT-PCR as previously described.⁶³ Because successful colonization was required for establishment of the infection model, only P. gingivalis-positive mice were included in subsequent analyses. Final group numbers were 5 mice in the P. gingivalis group and 6 mice in the P. gingivalis + 4-NQO group. At week 34, mice were euthanized, and the esophagus was opened longitudinally, photographed, and fixed in 4% paraformaldehyde. Gross morphology, wall thickness, and papillomatous lesions were documented. Lesions and adjacent esophageal mucosa were collected, homogenized in RIPA lysis buffer containing phosphatase and protease inhibitors, and centrifuged. Protein concentrations were determined by BCA assay. Equal amounts of protein were separated by SDS-PAGE, transferred to PVDF membranes, and immunoblotted with antibodies against p-STAT3 (Tyr705) and total STAT3. Bands were visualized by enhanced chemiluminescence and quantified by densitometry. Cervical lymph nodes (CLNs) were collected and placed in 6-well plates containing 3 mL of ice-cold PBS. Tissues were mechanically dissociated using frosted slides, followed by enzymatic digestion (1 mg/mL collagenase A, DNase, and 2% heat-inactivated FBS in PBS) for 20 min at 37°C with 5% CO₂. The cell suspension was passed through a 70-μm strainer, treated with 0.1 M EDTA for 5 min at 37°C, and washed twice with PBS. After red blood cell lysis using ammonium chloride potassium buffer, cells were collected by centrifugation (450 × g, 10 min, 4°C) for subsequent staining and flow cytometry analysis.

Method details

Specimen collection and storage

To minimize potential cross-site contamination, oral samples were consistently collected prior to endoscopy under standardized conditions. All subjects fasted for at least 8 h and refrained from oral hygiene activities (toothbrushing, mouthwash, or flossing) for 12 h. Oral microbiome samples were collected in the morning using iCleanhcy flocked swabs (Shenzhen HuaChenYang Technology Co., Ltd., cat. #CY-98000-P1). Each swab was gently rubbed along the gingival margins of the second and third molars in all four quadrants, with each quadrant swabbed five times to ensure consistent sampling of both subgingival and supragingival sites. Immediately after collection, swab heads were placed into sterile tubes containing 1 mL of nucleic acid preservative solution, kept on ice, and transferred to −80°C within 30 min. Paired esophageal tissue specimens were collected using sterile biopsy forceps during endoscopy, following standard clinical protocols designed to minimize contamination. Tissues were snap-frozen on dry ice immediately after excision and stored at −80°C until analysis. Care was taken to avoid contact with non-target mucosal surfaces during sampling, and all specimens were handled using sterile techniques to reduce potential oral microbial carryover. All specimens were shipped on dry ice under continuous temperature monitoring.

16S rRNA gene amplification, sequencing, and bioinformatics

DNA extraction was performed using standard protocols optimized for high-yield, broad-spectrum recovery.⁶⁴ High-depth 16S rRNA gene sequencing was carried out. Five hypervariable regions of the 16S rRNA gene (V2, V3, V5, V6, and V8) were amplified using independent primer pairs targeting distinct regions of 16S rRNA gene, with primer sequences listed in Table S2. Amplification reactions were performed separately for each primer set, and the resulting amplicons were pooled prior to library preparation and sequencing on the NovaSeq 6000 platform. This multi-region approach improves taxonomic resolution and coverage while reducing amplification bias, thereby enhancing species-level classification and detection sensitivity, particularly in low-biomass tumor samples.^65,66 Raw reads were demultiplexed and quality filtered (adapter removal using cutadapt v1.9; merging using FLASH v1.2.8; trimming with fqtrim v0.94 for Q < 20 and length <100 bp; and chimera removal using Vsearch v2.3.4), denoised (DADA2 within QIIME2), and assigned to amplicon sequence variants (ASVs) using the SILVA and NT-16S reference databases with the QIIME2 feature-classifier. The Short MUltiple Regions Framework (SMURF) was applied for region-integrative taxonomy profiling, referencing GreenGenes with contamination filtering. Alpha and beta diversity analyses were performed in QIIME2. Relative taxonomic abundance, Wilcoxon tests for differentially abundant genera, LEfSe analysis (LDA ≥ 3.0, p < 0.05, nsegata-lefse), and data visualization were conducted using R (v3.4.4) in the OmicStudio platform (https://www.omicstudio.cn/tool) using standardized analysis pipelines.⁶⁷ For diagnostic modeling, receiver operating characteristic (ROC) curve analysis and calculation of the area under the curve (AUC) for individual species or combinations of species were performed online using MetaboAnalyst 5.0 based on a logistic regression model. The AUC was calculated according to the following formula:

$\text{AUC}=\sum _{i=1}^{n-1}\frac{1}{2}\times \left(FP{R}_{i+1}-FP{R}_i\right)\times \frac{\left(TP{R}_{i+1}+TP{R}_i\right)}{2}$, where n represents the number of points on the ROC curve, and FPR and TPR denote the false positive rate and true positive rate at the i-th point, respectively. AUC values closer to 1 indicate higher diagnostic performance.

PICRUSt functional prediction analysis

Functional prediction of microbial gene content in oral swab and esophageal tumor samples was performed using PICRUSt2 (Phylogenetic Investigation of Communities by Reconstruction of Unobserved States version 2 (v.2.3.0) with software package: https://github.com/picrust/picrust2). Quality-filtered 16S rRNA amplicon sequence variant (ASV) tables generated by QIIME2 (see above), with taxonomy assigned at genus or species level, were used as input to infer the putative metagenomic profiles for each sample. Predicted gene family abundances were mapped to the Kyoto Encyclopedia of Genes and Genomes (KEGG) Orthology database to enable functional annotation and pathway-level profiling. KEGG pathways were collapsed at the pathway level, with emphasis on microbial-relevant functional categories, including metabolism-, transport-, and glycan-related pathways. To improve biological interpretability and avoid mis-annotation, predicted pathways were curated to exclude or annotate features considered ambiguous, non-bacterial (e.g., eukaryote-specific), or likely arising from database annotation overlap. This curation was guided by prior literature and established microbiome functional studies. Additional filtering criteria and thresholds used during pathway curation are described to improve transparency and reproducibility. Specifically, bioinformatic analyses were performed using the OmicStudio platform (https://www.omicstudio.cn/tool) with standardized pipelines. Differential pathway abundance between ESCC and healthy control groups was assessed using Wilcoxon rank-sum test. Pathways were ranked by statistical significance, and the top 30 pathways (smallest p-values) were selected for downstream visualization and interpretation. All functional predictions were interpreted with consideration of the inherent limitations of 16S rRNA-based inference, including reduced resolution and potential bias in gene content prediction. Accordingly, analyses focused on pathways plausibly associated with functional potential, and results are presented as exploratory and hypothesis-generating rather than definitive functional evidence.

Cell staining and flow cytometry

Lymphocytes isolated from CLNs were washed twice with Flow Cytometry Staining Buffer. Cells were stained with Zombie Aqua dye to exclude dead cells and incubated with CD16/CD32 antibody (ThermoFisher, cat. #14-0161-82) for 20 min on ice to block nonspecific Fc receptor binding. Following Fc block, cells were stained with fluorescently labeled anti-mouse antibodies: CD11c-PerCP-Cy5.5 and PD-L1-PE (ThermoFisher, cat. #12-5982-82), or matched isotype controls, Armenian Hamster IgG (PerCP/Cy5.5, cat. #45-4888-80) and Rat IgG2a kappa (PE, cat. #12-4321-80), for 30 min at 4°C in the dark. After staining, cells were washed with staining buffer and resuspended in PBS. Samples were acquired using an LSRFortessa flow cytometer (BD Biosciences), and data were analyzed with FlowJo v10 software (BD Biosciences).

Immunohistochemistry (IHC) and quantitative PCR

Fixed tissues were paraffin-embedded, sectioned, and stained with hematoxylin & eosin (H&E) for evaluation of tumor incidence, dysplasia, and tissue invasion. For IHC, sections were probed with anti-phospho-STAT3 (pSTAT3) antibody (Cell Signaling Technology, cat. #9145S), followed by incubation with secondary antibodies and visualization with DAB chromogen as previously described.^19,31 Slides were counterstained, scanned, and scored by independent, blinded, investigators for pSTAT3 positivity and intensity. Total RNA was extracted from mouse esophageal tissues using TRIzol reagent (Invitrogen, cat. #15596018), reverse-transcribed, and subjected to quantitative PCR (qPCR) for P. gingivalis 16S rRNA and cytokine mRNAs (IL-6, TNF-α, and IFN-γ). Primer sequences are provided in Table S3. Standard curves and triplicate reactions were used for quantification, and expression levels were normalized to GAPDH.

Quantification and statistical analysis

Categorical variables (e.g., sex, smoking) were compared via Pearson’s chi-square or Fisher’s exact test; continuous variables with normal distribution using Student’s t test (e.g., age); and non-normal distributed variables using Wilcoxon rank-sum test (e.g.,: alpha diversity indices, relative abundance at the genus-species level, and correlative analyses). Beta diversity was assessed by PCoA based on Weighted UniFrac and Bray-Curtis, with ANOSIM to evaluate group significance. Weighted UniFrac distance was applied to oral swab samples to incorporate phylogenetic relationships, whereas Bray-Curtis distance was used for esophageal tissue samples, as 5-region (5R) 16S sequencing does not support reliable phylogenetic tree construction required for UniFrac-based analysis. LEfSe analysis was used to identify taxonomic biomarkers (LDA>3.0, p < 0.05). Microbial co-occurrence networks were constructed using Spearman’s correlation (|ρ| > 0.4, p < 0.05). Diagnostic performance of ESCC microbial markers was evaluated using receiver operating characteristic (ROC) curves and the area under the curve (AUC). All analyses used SPSS v26.0 and R v3.4.4; p < 0.05 was considered statistically significant. For differential abundance analyses, multiple-testing correction was applied using the false discovery rate (FDR) method, and adjusted q-values are reported where applicable (Tables 2, 3, and 4). For ROC-based biomarker evaluation, candidate taxa were pre-selected using LEfSe analysis, and ROC curves were subsequently generated for these features; therefore, additional multiple-testing correction was not applied at this stage.

Additional resources

Additional sequencing data have been deposited in the Genome Sequence Archive (GSA) database at the National Genomics Data Center under accession numbers CRA036460 (https://ngdc.cncb.ac.cn/gsa/browse/CRA036460), and CRA036462 (https://ngdc.cncb.ac.cn/gsa/browse/CRA036462), and are publicly available.