A generalizable cross-continent prediction of esophageal squamous cell carcinoma using the oral microbiome

Abstract

Background

Esophageal squamous cell carcinoma (ESCC) has a poor prognosis and limited tools for early detection. Saliva is easily accessible and its microbiome composition can serve as a marker for upper gastrointestinal tract disease. This study aims to evaluate the potential of an oral microbiome signature for classifying ESCC.

Methods

In a cross-sectional study of 48 ESCC patients and 110 controls from South Africa, a region with high ESCC incidence, we studied the potential utility of an oral microbiome signature for the disease. We built models using nested cross-validation to evaluate whether this signature is generalizable to held-out samples and further evaluated generalizability in studies from China, a distinct geographic region.

Results

We find significant alterations in the oral microbiome in patients with ESCC including significantly reduced α diversity and increased abundance of Fusobacterium nucleatum. We also find that logistic regression models based on microbiome data can better classify ESCC in held-out samples (auROC=0.96) compared to clinical and demographic data (auROC = 0.69; DeLong p < 1 x 10^-8). Lastly, we find that microbiome-based models trained across multiple studies can generalize well to geographically distinct studies.

Conclusions

Our results show that the oral microbiome in individuals with ESCC is distinct from controls and that this signal can generalize across unseen samples, suggesting the potential of saliva to serve as a non-invasive screening tool for ESCC.

Plain language summary

Esophageal squamous cell carcinoma (ESCC) is a type of cancer that starts in the cells lining the esophagus, which connects the throat and stomach. It has a poor prognosis, and there is a need for non-invasive diagnostic methods. Saliva is easily accessible and the abundances of microbes in the oral cavity (the “oral microbiome”) can potentially be used to indicate the presence of cancer in the esophagus. This study explores the ability of computational models to use oral microbiome data to identify the presence of ESCC. Our results demonstrate that individuals with ESCC have a noticeably different oral microbiome compared to healthy individuals. We also show that classification models can distinguish ESCC across different geographical regions. These findings pave the way for the development of non-invasive tools for diagnosing ESCC.

ElNaggar et al. evaluate the potential of the oral microbiome to classify esophageal squamous cell carcinoma (ESCC). They find that there is a distinct oral microbiome signature in patients with ESCC, and that oral microbiome-based models can classify ESCC across geographically distinct studies.

Introduction

The dominant histological subtype of esophageal cancer globally is squamous cell carcinoma (ESCC), which has a very poor prognosis and remarkably high incidence rates in Eastern Asia and Eastern and Southern Africa, as opposed to esophageal adenocarcinoma (EAC), which predominates in Western countries¹. Early detection of ESCC is a major clinical challenge. The lack of early symptoms and the inaccessibility of endoscopic screening in high-risk areas result in late-stage diagnoses and high mortality rates². Established ESCC risk factors, including tobacco and alcohol use, polycyclic aromatic hydrocarbon exposure, hot food and beverage consumption, and poor oral health, can promote carcinogenesis^3,4. However, these risk factors only partially explain the extremely high incidence of this disease in certain regions of the world and have not informed screening practices.

Increasing evidence suggests that the esophageal microbiome, which is in direct contact with the esophageal mucosa, may modulate the risk of epithelial cancers by way of immune activation and chronic inflammation⁵, and may influence ESCC treatment response^6,7. Additionally, previous studies have reported alterations of the esophageal and oral microbiome in ESCC^8,9. Such associations may pave the way for the development of a microbiome-based biomarker for the detection of early-stage ESCC, which could potentially improve patient outcomes¹⁰. Our group previously showed that the salivary microbiome can distinguish patients with advanced precancerous changes and early esophageal adenocarcinoma¹¹, suggesting that the salivary microbiome may also be useful for the identification of ESCC patients at early, treatable stages.

Given the known variation in the human microbiome across different populations and regions in the world, identifying robust associations between the microbiome and ESCC across various populations is essential for the development of a microbiome-based diagnostic. In this study, we focus on the oral microbiome of an understudied population in South Africa, where there is a high incidence of ESCC^12,13. Several studies from China, which has regions of high ESCC incidence, have reported differences distinguishing the oral microbiome of ESCC patients^8,14–19. However, there remains a need to determine whether oral microbiome differences exist in other high-incidence regions of the world and whether they generalize across populations.

While direct sampling of the esophageal microbiome involves invasive procedures such as upper endoscopy, the sampling of saliva is relatively straightforward, and evidence shows that the salivary microbiome is strongly associated with the esophageal microbiome²⁰. For this reason, the salivary microbiome may serve as a “window” to the esophageal microbiome and associated diseases. Additionally, some studies show that salivary microbiome composition is temporally stable within individuals, suggesting its potential to provide diagnostic information regardless of sampling time, although some of the evidence is mixed in this regard^21–23.

Here, we investigate whether the oral microbiome is a reliable marker of ESCC in a case-control study of 158 individuals from a high-incidence region in South Africa, including 48 ESCC patients and 110 matched controls. We find that the oral microbiome is significantly associated with ESCC and identify specific microbes, including Fusobacterium nucleatum, Lautropia mirabilis, Veillonella dispar, and Prevotella salivae, with higher abundance in individuals with ESCC. We then evaluate the generalizability of oral microbiome-based models in identifying ESCC patients across four studies, including ours and three others conducted in China, and show that these models can identify individuals with ESCC from held-out studies. Overall, we find that compared to baseline clinical factors, the oral microbiome accurately classifies ESCC.

Methods

Study design

This cross-sectional study was designed to assess differences in oral microbiome composition between ESCC cases and matched controls. From September 2021 to November 2022, 55 Black South Africans with histologically confirmed ESCC were recruited across two GI clinics at the Chris Hani Baragwanath Academic Hospital in Soweto and the Charlotte Maxeke Johannesburg Academic Hospital in Johannesburg. Although patients did not undergo cancer staging, many individuals presented with dysphagia (trouble swallowing) and weight loss. Histological confirmation was obtained for all cancer cases. We excluded seven enrolled cases from our salivary microbiome analysis: histologic subtype could not be confirmed for 4 participants, one participant had EAC, one participant had gastric cancer, and one participant did not show evidence of cancer on histological assessment. Controls were enrolled 2:1 to ESCC cases from the Soweto area via PROMISE-SA, a multiple myeloma population-based screening study, between January 2022 and March 2023, and were frequency-matched to cases by age (± 5 years), sex, self-reported race, and study site. Matching by age was limited by the relatively small number of older eligible controls, resulting in a slight imbalance in age between cases and controls. It was confirmed via questionnaire that controls were between the ages of 40 and 75, self-identified as being Black or of African descent, had no current diagnosis of any cancer, and had no symptoms of dysphagia. Neither cases nor controls had a previous diagnosis of head or neck cancer, or any other type of cancer. At the time of saliva collection, none of the patients had received treatment for ESCC, including chemotherapy, radiation, or surgery.

We collected demographic and clinical data from each participant. Additionally, we collected 2 mL of saliva, without the use of a saliva stimulant, using the Oragene OG-500 DNA Saliva Collection Kit (DNAGenotek, Ontario, Canada). In our microbiome analyses and predictive modeling, we included the following clinical variables due to their potential relevance to ESCC and/or salivary microbiome composition: patient age (years), sex, marriage status, location of residence, cooking location, highest level of education completed, smoking, alcohol consumption, hot beverage consumption HIV status, and history of diabetes mellitus (Table S1). Clinical and demographic data were unavailable for two controls. Additional data is missing for age (3 ESCC), sex (2 ESCC), location of residence (2 ESCC; 1 control), cooking location (3 ESCC), education (3 ESCC), smoking (2 ESCC), alcohol consumption (2 ESCC; 1 control), hot beverage consumption (5 ESCC), diabetes status (2 ESCC), and HIV status (4 ESCC; 21 controls). In Table S1 only, education levels Grade 1 through 7 are collapsed into “Primary school” and Grade 8 through 12 are collapsed into “Secondary school.”

All enrolled patients provided written consent. The study was conducted in accordance with the principles of the Declaration of Helsinki and approved by the Human Research Ethics Committee at the University of the Witwatersrand (certificate number: M180306) and by the Institutional Review Board at Columbia University.

Statistics and reproducibility

Sample size calculations were made based on differences in α diversity between cases and controls; a projected sample size of 50 cases and 100 controls provided 82% power to detect a 0.5 standard deviation difference, assuming a type I error rate of 0.05. Between-group comparisons of demographic variables in Table S1 were performed using chi-squared tests for categorical variables and Mann-Whitney U tests for continuous variables. Unadjusted α-diversity comparisons between cases and controls were conducted using Mann-Whitney U tests, and adjusted comparisons were conducted using a logistic regression model that included the clinical and demographic variables listed above. β-diversity comparisons were conducted using permutational multivariate analysis of variance (PERMANOVA) with 10,000 permutations to calculate p values. Adjusted PERMANOVA tests included the clinical and demographic variables listed above in the model. The DeLong test was used to assess the statistical significance of differences in machine learning model performance and Mann-Whitney U tests were used to assess whether models performed better than a random classifier. Statistical significance was defined as p < 0.05. In the case of multiple comparisons, reported p values are corrected for false discovery rates (FDR) using the Benjamini-Hochberg method²⁴.

Microbiome sequencing and analysis

Microbial DNA was isolated in 11 separate batches from saliva specimens using the QIAamp BiOstic Bacteremia DNA kit as per the manufacturer’s protocol. Negative extraction blanks were processed in tandem for 9 out of 11 batches. The V1-V2 region of the 16S rRNA gene was amplified using Illumina adapter-ligated primers (27F-338R). The forward primer used was AGAGTTTGATCCTGGCTCAG and the reverse primer used was TGCTGCCTCCCGTAGGAGT. The resulting libraries were barcoded and sequenced using the Illumina MiSeq platform.

To limit the analysis to bacterial reads only, reads mapped with bowtie2 (ref. ²⁵) with inclusive parameters to CHM13-V2 human and PhiX genome sequences were removed from further analysis. The QIIME 2 v2024.2 microbiome analysis platform, which is a wrapper for various microbiome bioinformatics tools, was used for the following analyses²⁶. DADA2 v1.26.0 was used to filter low-quality reads, merge paired-end reads, and identify a total of 11,277 amplicon sequence variants (ASVs) and their counts in each sample²⁷ (Table S2). A median of 44,788 clean and non-chimeric reads were available per sample for analysis. All ASVs were aligned with mafft²⁸ and used to construct a phylogeny with fasttree2 (ref. ²⁹). Taxonomies were assigned to ASVs using the q2-feature-classifier classify-sklearn naive Bayes taxonomy classifier³⁰. A classifier trained on Human Oral Microbiome Database v15.23 (HOMD) OTU sequences was used to assign taxonomies to each ASV for interpretability³¹ (Table S3), except in cross-study species-level analyses where we used Greengenes2 2022.10 (detailed below).

Possible bacterial contaminants were removed in silico using SCRuB, which probabilistically estimates the true counts in a sample based on the proportions of taxa observed in the negative extraction blanks³². Nine out of eleven batches were “decontaminated” using their corresponding blanks, while the two batches without corresponding blanks were decontaminated using a pooled blank which summed taxa across all available blanks.

Sample α diversity was estimated using the Shannon index, and inter-sample diversity (β diversity) was estimated using unweighted and weighted UniFrac³³. Prior to diversity estimates, rare taxa were filtered (only taxa present in at least 5 samples were included) and each sample was rarefied to 9700 reads, which was the minimum read depth across samples. 2091 out of 11,277 ASVs were retained post-filtering. Community-level differences between ESCC samples and control samples were tested via PERMANOVA using the “adonis” function within the QIIME2 q2-diversity plugin, once without adjusting for covariates and once adjusting for all available clinical and demographic covariates in the model, including age, sex, location of residence, education, marriage status, hot beverage consumption, smoking, diabetes, HIV, cooking location, and batch³⁴.

Differential abundance testing

Differential abundance analysis was performed in R v4.3.2 using beta-binomial regression models from Corncob v.4.1.³⁵. ASV tables were filtered to include taxa present in five or more samples, which resulted in 2,091 ASVs retained. Models were adjusted for batch, age, alcohol consumption, cooking location, hot beverage consumption, and smoking; corncob adjusts for sequencing depth in its model by default. To identify differentially abundant taxa, we assessed which taxa had FDR-corrected p values of less than 0.1 in a model only adjusted for batch and then identified which of these taxa were also significant (FDR-corrected p < 0.05) in the full model, which included the rest of the covariates. For the Fusobacterium dominance analysis, models were additionally adjusted for the total abundance of the genus in each sample. Reported p values are FDR corrected using the Benjamini-Hochberg method²⁴; taxa with significant p values are displayed in Tables S4 (genus), S5 (ASV), and S6 (Fusobacterium ASVs).

Training, testing, and evaluation of ESCC classifiers

Supervised prediction models were built to classify ESCC versus control samples in our cohort using the scikit-learn Python library (v1.3.1). Logistic regression was used for all classification tasks due to its interpretability and comparable performance to more complex models, including lightGBM, support vector machine, and random forest (Supplementary Fig. 1). Model performance was evaluated across four feature sets: (1) clinical and demographic variables (including age, sex, marriage status, education, residence location, cooking location, smoking, alcohol consumption, hot beverage consumption, HIV status, and diabetes status); (2) microbiome data at the ASV level (11,277 features); (3) microbiome data at the species level (Greengenes2; 576 features); and (4) a combined dataset consisting of both clinical and microbiome (ASV) data (i.e., 1 + 2).

For all models, samples were split up into training and test sets by grouping samples by batch and creating eleven test-train splits, leaving one batch as a held-out test set each time (i.e. eleven-fold cross-validation). This was done to account for confounding effects in our models introduced by class imbalances in each batch (e.g., some batches consisted only of ESCC samples). For the clinical models, missing information was imputed using the median of the training set for continuous variables and the mode of the training set for categorical variables. To tune hyperparameters, we used nested cross-validation. Model hyperparameters in our microbiome models included steps for pre-processing and feature selection (Table S7). All microbiome features with a variance of 0 were removed and the remaining features were centered log-ratio (CLR) transformed³⁶. Feature selection was implemented as part of the nested cross-validation procedure to ensure that feature selection was performed only on the training data in each fold, thereby preventing information leakage in the validation or test set. We fit a lasso model to the data and then retained all features with non-zero coefficients. We tuned a hyperparameter k, which controls the proportion of features to retain. For each value of k, we determined the optimal L1 regularization strength (α) that results in approximately k percent of the input features having nonzero coefficients. Each training set was split into five folds (i.e., ‘inner folds’) on which we used 1,000 iterations of a random set of hyperparameters. This process was repeated five times to account for stochasticity. The best hyperparameter set was selected as the model with the highest average area under the receiver operating characteristic curve (auROC) score based on performance on the inner folds. This model was then trained on the entire training data and evaluated once on the unseen held-out batch. For model evaluation, we calculated the overall auROC and area under the precision recall curve (auPR) across all folds together as well as the per-outer-fold auROC and auPR statistics. The mean and standard deviation across folds were computed for 9 of 11 outer folds in nested cross-validation; two folds containing only controls were excluded from per-fold calculations.

Cross-study preprocessing and validation of ESCC classifiers

We conducted a literature search to identify studies suitable for external validation of our microbiome-based ESCC predictor. 12 studies were initially identified that sampled the oral and/or esophageal microbiome of individuals with ESCC and non-ESCC controls^{8,14–19,37–41}. We excluded studies without publicly available 16S rRNA sequencing data of saliva samples.

We reprocessed PRJNA660092¹⁸, PRJNA587078¹⁵, and PRJNA961904³⁷ to the ASV-level using the same steps detailed above. For all three cohorts, including our own, we assigned taxonomies using the q2-feature-classifier classify-sklearn naive Bayes taxonomy classifier trained on the Greengenes2 2022.10 database, with the ‘–p-confidence’ parameter set to 0. This approach ensured species-level assignments for all ASVs, regardless of confidence scores, to enable a consistent species-level harmonization across all studies. We filtered out features that appeared in less than 10% of samples, which resulted in 315 features being retained, and CLR-transformed the remaining features. Then, we performed leave-one-study-out cross-validation, where we trained a logistic regression with no penalty on all but one study, and evaluated the model’s performance on each held-out study. For the species-level model trained on our cohort and tested on the external cohorts, we retrained each nested model on the entire South Africa cohort and individually tested each of the eleven models on the external cohorts. For each external cohort, we report the mean and standard deviation of the auROC and auPR across all eleven models. Lastly, we performed 10-fold nested cross-validation within each external study using species-level microbiome data. We report the overall auROC and auPR as well as the mean and standard deviation of the auROC and auPR across folds. Hyperparameter optimization and preprocessing steps on the external studies were identical to those used on our data. To assess whether models performed better than a random classifier, we conducted a one-sided Mann-Whitney U test to evaluate whether predicted scores were higher for the positive than for the negative class. Predicted values for all models are included in Table S8.

Results

Patient characteristics from a South African ESCC case-control study

We enrolled 55 adult patients in Soweto, Johannesburg with histologically confirmed ESCC. We additionally enrolled 110 controls with no history of any cancer and no dysphagia symptoms. At the time of sample collection, none of the patients had received treatment for ESCC, including chemotherapy, radiation, or surgery. We were unable to obtain histologic confirmation of ESCC for 7 cases and excluded these from analyses. Controls were geographically co-located, and all samples were processed together. We frequency matched controls to cases, to the extent possible, by age (± 5 years), sex, and location, and characterized their salivary microbiomes using 16S rRNA gene amplicon sequencing (Methods). Patients with ESCC were older (two-sided Mann-Whitney U test p < 0.001), and less likely to drink alcohol (Chi-squared p = 0.029), drink hot beverages (p = 0.010), and cook outside (p = 0.030; Table S1). There were no significant differences between ESCC and controls with regards to sex, marriage status, place of residence, smoking, self-reported HIV status, and diabetes status (Table S1).

The oral microbiome in ESCC is distinguishable from controls

To investigate overall microbiome differences between ESCC patients and controls, we first tested whether within-sample diversity (α diversity) varied between groups. ESCC samples had significantly lower α diversity (two-sided Mann-Whitney U test p = 2×10⁻⁴ and p = 10⁻⁸ for Shannon and Chao1, respectively; Fig. 1aand Supplementary Fig. 2a). We observed no differences in Shannon α diversity based on age, sex, smoking, HIV status, or extraction batch (p > 0.2 for all), although drinking hot beverages was associated with higher α diversity (p = 0.027). Shannon diversity remained significantly associated with ESCC in a logistic regression adjusting for all available clinical and demographic covariates in addition to experimental batch (p = 0.009). Additionally, we found that the oral microbiomes of ESCC patients clustered separately from those of controls when considering only the presence and absence of amplicon sequence variants (ASVs) (unweighted UniFrac, PERMANOVA p < 0.0001; Fig. 1b), and this separation remained distinct even when adjusting for all available covariates (p < 0.001). However, ESCC and controls did not cluster separately when microbiome distances were weighted by abundance (weighted UniFrac, p = 0.096; Supplementary Fig. 2b). This suggests that low-abundance ASVs may be contributing to differences between ESCC and controls, as opposed to community-level differences. To assess the contribution of low-abundance ASVs to differences between ESCC and controls, we subset the data to include only ASVs in the lowest percentiles of mean abundance and evaluated the relationship between explained variance by case-control status (adonis R²) and the percentile threshold (Supplementary Fig. 2f). The lowest abundance ASVs contributed to the greatest separation between cases and controls, and explained group variance decreased as higher abundance ASVs were progressively included (Supplementary Fig. 2c-e).

Fig. 1. The oral microbiome is associated with ESCC.

A Box and swarm plots (Box, IQR; line, median; whiskers, 1.5xIQR) showing significantly lower α diversity in ESCC patients (N = 48) compared to controls (N = 110; two-sided Mann-Whitney U test p = 2 × 10⁻⁴). B PCoA of unweighted UniFrac distances demonstrated significant clustering of patients with ESCC (PERMANOVA p = 10⁻⁴). Ellipses represent 2 standard deviations.

Genus and ASV-level abundances are associated with ESCC

We next assessed whether specific genera were significantly associated with ESCC, with and without adjustment for covariates (Methods; see Table S4 for unadjusted results). We found that Capnocytophaga, Lautropia, Arachnia, Streptococcus, Selenomonas, Leptotrichia, and Campylobacter were significantly elevated in the saliva of ESCC patients (FDR-corrected p < 0.05 for all with covariate adjustment). Additionally, Filifactor and Bacteroides were depleted in the ESCC oral microbiome (FDR-corrected p = 0.003 and p = 0.01, respectively; Fig. 2a, Table S4).

Fig. 2. Specific microbial taxa are associated with ESCC.

A–C Box and swarm plots (line, median; box, IQR; whiskers, nearest point to 1.5*IQR) showing the relative abundances (log scale) of differentially abundant genera (A), ASVs (B) and ASVs relative to the genus Fusobacterium (C) in the oral microbiome between ESCC and controls. Beta-binomial regression p values calculated using two-sided parametric bootstrap Wald tests in corncob³⁵, with FDR-correction (Benjamini-Hochberg) are displayed. Blue boxes represent controls (N = 110) and red boxes represent ESCC cases (N = 48).

Analyzing differential abundance of bacteria at the ASV level can offer more detailed insight than genus-level analyses, as microbial functions are often species-specific. Therefore, we investigated associations between ESCC and individual ASVs while adjusting for covariates (Methods). A total of five ASVs were identified as differentially abundant (FDR-corrected p < 0.05 for all): ASV4873 (annotated as P. salivae), ASV3816 (V. dispar), ASV3722 (L. mirabilis), and ASV5088 (F. nucleatum) were elevated in the ESCC oral microbiome, and ASV7524 (Absconditabacteria sp.) was depleted in ESCC (Fig. 2b, Table S5).

F. nucleatum is an oral commensal that is thought to promote the development of ESCC, is present in high abundance in a subset of ESCC cases, and is associated with worse clinical outcomes^42,43. As we showed, higher abundance of oral ASV5088 (F. nucleatum) was associated with the ESCC oral microbiome. We further assessed the ability of ASV5088 to differentiate individuals with ESCC and calculated an auROC of 0.73, highlighting its discriminative ability (Supplementary Fig. 3). However, we did not identify genus-level Fusobacterium abundances associated with ESCC. We therefore next investigated whether certain ASVs were more likely to exist in higher proportions in the Fusobacterium genus between ESCC and controls, i.e., whether any ASVs were more prevalent relative to others within the genus. We found that ASV6145 (Fusobacterium periodonticum) was significantly decreased in ESCC relative to other Fusobacterium ASVs (p = 0.003; adjusted for genus abundance in sample and ESCC risk factors). Additionally, ASV5088 (F. nucleatum) was significantly increased in ESCC relative to other Fusobacterium species (p = 0.01) (Fig. 2c, Table S6). The dominance of F. nucleatum within the Fusobacterium genus in ESCC cases recapitulates its potential role in the disease.

The oral microbiome accurately classifies ESCC

The ability of the oral microbiome to classify ESCC status holds great potential for the development of a non-invasive diagnostic. We therefore devised logistic regression models and checked whether the oral microbiome can be used to distinguish ESCC patients from controls. As a benchmark, we also tested models based on all available clinical information. Because the samples were processed in batches, we evaluated models on held-out batches (leave-one-batch-out cross-validation) to limit possible confounding effects^44,45 (Methods). Model hyperparameters were tuned using the training set (“nested” cross-validation) without information leakage from the test set.

Models using clinical and demographic information obtained moderate accuracy with an overall auROC of 0.69 (mean ± std across folds of 0.72 ± 0.16) and an overall area under the precision-recall curve (auPR) of 0.54 (0.67 ± 0.2; Fig. 3a; Supplementary Fig. 4a). Using oral microbiome data at the ASV level, we were able to generate a model with significantly higher accuracy (overall auROC = 0.96, mean ± std 0.95 ± 0.046 across folds; auPR = 0.92, 0.91 ± 0.11; DeLong’s test p = 9.12 × 10^–9 vs. model using clinical data). We also built a species-level model, which performed slightly worse compared to the ASV-level model (overall auROC = 0.92, folds 0.93 ± 0.075; auPR = 0.85, 0.92 ± 0.09; p = 0.046 vs. ASV-level model). A model combining clinical and ASV-level oral microbiome data level did not improve upon the model based solely on oral microbiome data (overall auROC = 0.96, folds 0.95 ± 0.058; auPR = 0.93, 0.92 ± 0.09; p = 0.72 vs. ASV-level model). This suggests that oral microbiome composition may be able to classify ESCC with high accuracy, and that its information content encompasses that of the associated clinical and demographic characteristics.

Fig. 3. Accurate microbiome-based classification of ESCC.

A Receiver operating characteristic (ROC) curves comparing ESCC classification accuracy for ASV-level microbiome models (auROC = 0.96), species-level models (auROC = 0.92), models based on clinical data (auROC = 0.69), and models based on both microbiome (ASV) and clinical data (auROC = 0.96), evaluated on held-out experimental processing batches (Methods). B Effect on prediction (SHAP values) for the top ten most predictive ASVs in the ASV-level microbiome-based model, sorted by importance. Each dot represents a specific sample, with the color corresponding to the relative value of the ASV in the sample compared to all other samples. C ROC curves showing the performance of species-level microbiome predictors of ESCC, trained on our cohort (N = 158) and evaluated separately on three held-out studies from China: Zhao et al. 2020 (N = 91), Wang et al. 2019 (N = 41), and Chen et al. 2024 (N = 52). The model from each external cross-validation fold was evaluated separately, with the line showing mean ROC curves and shaded regions representing ±1 standard deviation. D ROC curves showing the performance of a species-level microbiome-based ESCC classifier, trained on all studies except one and evaluated on each held-out study.

We evaluated the importance of each feature towards the model prediction for each sample using SHapley Additive exPlanations (SHAP) values (Fig. 3b; Supplementary Fig. 5). An analysis of our clinical-based predictor showed that age was one of the most predictive features, which corresponds to the older age of the ESCC group (Supplementary Fig. 5a). Additionally, of the top ten most predictive features identified in our microbiome-based predictor, ASV5088 (F. nucleatum) was the only taxon also highlighted as significantly differentially abundant in ESCC patients (Supplementary Fig. 5b and Fig. 2). Interestingly, some taxa within the same genus had strongly opposing signals: for example, four species of Veillonella come up in model evaluation, two of which were associated with ESCC (ASV1169: Veillonella atypica and ASV10234: Veillonella parvula) and two other unidentified ASVs from this genus were associated with controls (ASV582 and ASV11255). Additionally, ASV505 (Streptococcus salivarius) was predictive of ESCC, while ASV1707 (Streptococcus parasanguinis) was associated with controls. These patterns indicate the importance of ASV-level analyses of oral microbiota in ESCC, as different microbes within a genus may play different roles in the oral cavity and in disease.

Oral microbiome-based models for ESCC generalize across studies

Several studies have been performed in high-incidence regions of China assessing relationships between the oral microbiome and ESCC^{8,14–19,37–41}. Of the studies we identified, three included publicly available salivary microbiome 16S rRNA sequencing data from ESCC patients and non-ESCC controls: a study including 41 individuals (20 ESCC; Wang et al. 2019), all with periodontitis or gingivitis (gum disease)^15,46; a study including 90 individuals (39 ESCC; Zhao et al. 2020)^18,47; and a study including 52 individuals, 31 of which had early-stage ESCC and 21 controls^37,48 (Chen et al. 2024). Wang et al. 2019 and Zhao et al. 2020 enrolled participants from the Henan, China region, albeit at different hospitals^15,18, and Chen et al. 2024 of the studies enrolled participants from Nanjing, China³⁷. All studies enrolled individuals who had not undergone previous treatment for ESCC, and their salivary microbiomes were profiled using the V3-V4 region of the 16S rRNA gene. Due to discrepancies in ASVs across these three studies and our study, which profiled the microbiome using the V1-V2 region, we were only able to evaluate our species-level microbiome-based model.

We first evaluated whether this model, trained on our cohort from South Africa, could independently generalize to each external cohort. We found that the model generalized well to Zhao et al. 2020 (overall auROC = 0.79, mean ± std across folds 0.79 ± 0.058; auPR = 0.77, 0.77 ± 0.07; p = 2.87×10⁻⁷ one-sided Mann-Whitney U test; Methods), although not as well to Wang et al. 2019, which enrolled patients with gum disease (overall auROC=0.56, folds 0.55 ± 0.059; auPR = 0.60, 0.60 ± 0.07; p = 0.16), or to Chen et al. 2024, which focused on patients with early-stage ESCC (overall auROC = 0.52, folds 0.52 ± 0.034; auPR = 0.65, 0.65 ± 0.023; p = 0.55; Fig. 3c, Supplementary Fig. 4b). To assess whether this poor generalizability was due to less pronounced separation between ESCC and controls in the external studies, we ran 10-fold nested cross-validation separately on each external study (Supplementary Fig. 6). We found that ESCC was highly distinguishable from controls using the species-level microbiome in Zhao et al. 2019¹⁸ (overall auROC=0.88, folds 0.94 ± 0.067; auPR = 0.86, 0.94 ± 0.063; p = 3.2 × 10⁻¹⁰) and was fairly distinguishable in the early-stage ESCC study³⁷ (overall auROC = 0.68, folds 0.64 ± 0.23; auPR = 0.76, 0.81 ± 0.13; p = 0.015). However, the oral microbiome was not able to distinguish ESCC well in the study consisting of only individuals with gum disease¹⁵ (overall auROC = 0.54, folds 0.54 ± 0.21; auPR = 0.50, 0.67 ± 0.15; p = 0.33), which would explain the poor generalizability of our model to this particular study.

Finally, we evaluated the potential for a more global microbiome-based diagnostic by checking whether models trained on all studies except one could generalize to the held-out cohort. In this leave-one-study-out cross-validation, oral microbiome-based models were able to identify ESCC in the held-out study with some accuracy (auROC = 0.64–0.81, auPR = 0.70–0.84; all p < 0.05; Fig. 3d, Supplementary Fig. 4c), indicating a potentially generalizable microbiome signature for ESCC across geographic regions. To investigate which taxa contributed most to predictions across studies, we identified the top 30 taxa with the largest mean logistic regression coefficients across the four models (Supplementary Fig. 7). Notably, these 30 taxa had coefficients in the same direction across all models.

Discussion

In this study, we characterized the oral microbiome of 158 individuals from South Africa using 16S rRNA gene sequencing, including 48 individuals diagnosed with ESCC and 110 healthy controls. We detected differences in the oral microbiome of ESCC compared to controls, including a significant decrease in microbial α diversity. We further found that genera including Capnocytophaga, Lautropia, Arachnia, Streptococcus, and Selenomonas were associated with ESCC, in addition to select ASVs including F. nucleatum, V. dispar, and L. mirabilis. Lastly, we demonstrated that microbiome-based models can classify ESCC status across geographically distinct cohorts, suggesting their potential as a screening tool for the disease. Prospective studies are necessary to evaluate whether the oral microbiome is unique prior to disease onset and if the microbiome can be used to classify the disease across cancer stages.

Many of the bacteria associated with ESCC identified in this study are known to promote inflammation and neoplasia in the oral cavity. Capnocytophaga, for example, has been demonstrated to invade oral squamous cell carcinoma cells and induce epithelial-to-mesenchymal transitions^49,50. Selenomonas species, commonly associated with periodontal disease, are known to attach to gingival epithelial cells and trigger inflammatory responses^51,52. Streptococcus, the most prevalent genus in the oral cavity, includes species such as S. mitis and S. anginosus that have been implicated in oral and upper gastrointestinal cancers^53–55. A study from Tanzania analyzed the ESCC tumor-associated microbiome and reported similar findings compared to what we found in the oral microbiome, including high abundances of Selenomonas, Streptococcus, and Campylobacter⁵⁶.

Our findings also support previous evidence implicating F. nucleatum in ESCC. Interestingly, a single ASV assigned to F. nucleatum (ASV5088) was able to discriminate well between ESCC and controls with an auROC of 0.73. Further validation in independent cohorts is needed to determine whether abundance of F. nucleatum, independently or in combination with other species or strains, can serve as a robust biomarker for ESCC using targeted assays such as qPCR. F. nucleatum has been shown to be enriched in ESCC tumor tissue, and evidence suggests that the microbe may invade ESCC cells and enhance cell growth, thereby aiding disease progression^40,42,43. Notably, we find that within the Fusobacterium genus, F. nucleatum tends to exist in higher proportions in ESCC compared to other species in the genera. By contrast, the oral microbiome of controls had a significantly higher proportion of F. periodonticum. Both F. nucleatum and F. periodonticum are known as active invader species, owing to their ability to independently invade epithelial host cells and, in the case of F. nucleatum, subvert host cell function^57–59. Based on this evidence, it is possible that F. nucleatum could be outcompeting F. periodonticum in the ESCC oral microbiome.

Previous studies have established a higher risk of ESCC in those with poor oral health, specifically tooth loss and lack of regular oral hygiene^60–63. Microbes involved in poor oral health may similarly be involved in inflammatory processes in ESCC, potentially explaining the link between poor oral health and cancer. A prospective study on esophageal adenocarcinoma (EAC) and ESCC found associations between Treponema forsythia and EAC as well as Porphyromonas gingivalis and ESCC⁹. These microbes are members of the “red complex” of periodontal pathogens, which are described as drivers of periodontitis⁶⁴. In our study, we identified no relationship between ESCC and any red-complex species. Although we did not have oral health data for our participants, previous research suggests that these oral microbiome alterations may exist independent of oral health status. In a study on the oral microbiome in patients with Barrett’s esophagus and early EAC¹¹, tooth loss was found to be associated with high-grade dysplasia and early EAC. However, even after adjusting for tooth loss, taxa associated with high-grade dysplasia and early EAC remained significant, indicating that the association of tooth loss with disease may be mediated through the oral microbiome. The causal pathway between the oral microbiome, oral health, and ESCC risk remains undefined.

Alcohol and hot beverage consumption are well-established as risk factors for ESCC, although we did not find these relationships reflected in our cohort from South Africa. On the contrary, hot beverage consumption and alcohol use were increased in controls. We speculate that this is because the relationship between ESCC and alcohol consumption is likely more complex and is contingent on the frequency of alcohol consumption and intake amount, information that was not captured in our study questionnaire. Similarly, the relationship between hot beverage consumption and ESCC may depend on temperature of the beverage, which also was not captured in our data. We also suspect that there may have been an element of reverse causation, whereby intake of alcohol and hot beverages exacerbated symptoms of ESCC in patients, leading to their avoidance.

A strength of our study is that we were able to devise accurate models for predicting ESCC using the oral microbiome in South Africa and across geographic regions, finding statistically significant signals even if not high accuracy (auROC of 0.64 for the least accurate model). Using cross-validation, we demonstrated that microbiome-based models were far superior to models using established clinical risk factors for the identification of patients with ESCC in South Africa. Additionally, we checked whether oral microbiome signatures of ESCC could generalize across diverse geographic regions. We evaluated species-level microbiome predictors on held-out studies, demonstrating that aggregate models generalized well to held-out studies, including one study that enrolled patients with early-stage ESCC. Species contributing to ESCC predictions are consistent across all aggregate models and reinforce our findings in South Africa-based analyses: Streptococcus, Capnocytophaga, and Veillonella species are associated with ESCC while F. periodonticum is consistently associated with controls. Lower performance of the individual South Africa-trained model on held-out studies from China can be explained by geographic effects, which are associated with the structure of the oral microbiome⁶⁵, as well as discrepancies in extraction kits and sequencing protocols which may also have introduced study-specific processing biases^66–68. Such differences are harder to capture for a model trained on one study with a lower sample size⁶⁹. However, relatively good performance of a simple aggregate model across distinct cohorts, including a study with early-stage ESCC patients, demonstrates that there exists a somewhat generalizable oral microbiome signature that would allow for the identification of individuals with ESCC independent of geographic region and potentially in earlier stages of the disease. This is promising for the development of microbiome-based diagnostics. In order to further establish geographic generalization, models should be assessed in other areas with high incidence of ESCC such as Iran and East Africa^1,70. Additionally, while evidence suggests that the salivary microbiome is highly stable over time compared to other body sites²¹, future work should specifically establish whether a salivary microbiome-based diagnostic for ESCC remains accurate regardless of sampling time.

The study had certain limitations. Although staging data was not available for each participant, most ESCC patients presented with dysphagia, likely due to the presence of larger tumors and late-stage disease. The oral microbiome may be somewhat distinct in patients with early and late-stage ESCC, although our models showed some generalizability in the leave-one-study-out cross-validation setting between an external study with early-stage ESCC and other studies with later-stage ESCC, including our own. Since microbiome-based models trained on the South Africa study alone did not generalize as well to an external study including individuals with early-stage ESCC or another external study including only individuals with gum disease, further work must be done to ensure that a salivary microbiome-based diagnostic can distinguish ESCC at earlier stages and not be influenced by the presence of periodontal disease. Due to a lack of publicly available clinical and demographic data associated with external studies, we could not compare the generalizability of a microbiome-based model versus a model based on clinical features. Additionally, participants with ESCC and controls should have been matched more closely by age and excluded based on factors such as antibiotic use, although this was not possible due to the limited availability of participants. As controls did not undergo upper endoscopy, we cannot confirm with certainty that they did not have ESCC. However, this was a relatively healthy cohort without dysphagia symptoms, and thus the likelihood of undetected ESCC in out controls is very low.

Overall, these results demonstrate the potential of the oral microbiome to distinguish ESCC from controls and identify specific bacterial ASVs implicated in the disease. Translation of these findings for the oral microbiome-based screening of ESCC in low-resource settings will require validation across distinct geographic regions, in larger sample sizes, and especially in patients with early, potentially curable ESCC. Such a test could be used to triage patients for upper endoscopy, which could have a major public health impact in very high-incidence regions.

Author contributions

J.A.A., W.C.C., C.G.M., A.I.N., and Z.J. conceived and designed the study. Recruitment and clinical assessment of participants were performed by W.C.C., L.M.P., T.K.M., and M.U.K. A.C.U. generated all data. S.E., Z.J., T.K., and J.A.A designed all analyses, conducted data analysis, and interpreted the results. All authors (S.E., W.C.C., L.M.P., T.K.M., M.U.K., C.G.M., P.R., Z.J., A.I.N., A.K.R., A.C.U., T.K., J.A.A.) contributed to writing and editing the manuscript and approved the final version. J.A.A. and T.K. supervised the study.

Peer review

Peer review information

Communications Medicine thanks David P. Kelsen and the other anonymous reviewer(s) for their contribution to the peer review of this work.

Data availability

The 16S rRNA gene amplicon sequencing data analyzed in this study is publicly available in the NCBI sequence read archive (SRA) under BioProject accession number PRJNA1271143. Data associated with external studies are publicly available under BioProject accession numbers PRJNA660092 (Zhao et al. 2020), PRJNA587078 (Wang et al. 2019), and PRJNA961904 (Chen et al. 2024). Numerical results underlying all figures are included in the Supplementary Data 1 file.

Code availability

Code to generate figures is available here https://github.com/korem-lab/ESCC_microbiome_16S_analysis (10.5281/zenodo.18319080).

Competing interests

Chen is an Editorial Board member for Communications Medicine but was not involved in the editorial review or peer review, nor in the decision to publish this article. Neugut has consulted for Otsuka, United Biosource Corp, Value Analytics, Merck, and Cybin, and has research funding from Otsuka and Kyowa Kirin. Abrams has consulted for Exact Sciences and Cyted Health and has research funding from Pentax Medical. All other authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s43856-026-01468-y.