The oral microbiome is associated with the diagnosis, prognosis and radiotherapy sensitivity of esophageal cancer Authors: Lei Li, Lingling Li, Yuanyuan Yang, Chaoguan Wang, Zhen Lan, Jingyuan Liu, Xiaohui Zhu, Chunyu Zhao, Mengzhao Yang, Yarong Ma, Yang Liu, Zhigang Ren Published: 2026 ABSTRACT BackgroundEsophageal cancer (EC) is a leading cause of cancer-related mortality worldwide and early detection strategies and precise postoperative interventions must be developed. However, the identification of noninvasive biomarkers for the diagnosis and prognosis remains limited.MethodsWe performed 16S rRNA gene sequencing on tongue-coating samples from 440 participants, including 157 EC patients, 167 healthy controls (HCs) and 120 EC patients who received radiotherapy. We characterized the oral microbiome and constructed microbial diagnostic and prognostic classifiers. Furthermore, the oral microbiome of EC who received radiotherapy (n = 120) was characterized.ResultsThe oral microbial diversity of EC patients was increased, with differences in the microbial community between EC patients and HCs. In EC, the genera Veillonella, Streptococcus and Actinomyces were enriched, whereas Porphyromonas and Rothia were depleted. The classifier based on six optimal microbial markers was constructed using random forest algorithm and achieved area under the curves (AUCs) of 93.69% and 95.18% in discovery and validation groups, respectively. After radiotherapy, the oral microbial diversity and richness were significantly reduced. The prevalence of opportunistic pathogens, including Fusobacterium and Porphyromonas, decreased, whereas the prevalence of Streptococcus and Actinomyces increased in EC patients after radiotherapy. Through six months of follow-up, patients were categorized into a progression group (PG) (n = 26) and a nonprogression group (NPG) (n = 106) based on the presence of tumor recurrence, metastasis, and death. A prognostic model based on 13 selected amplicon sequence variants (ASVs) of the oral microbiome was constructed, with an AUC of 99.29%. The random forest analysis identified six key differential ASVs between the PG and the NPG, including Fusobacterium and Gemella. Additionally, twenty-five ASVs associated with nine clinical indicators were identified.ConclusionsOur study provides a comprehensive characterization of the oral microbiome in both EC patients and EC patients after radiotherapy, highlighting the potential of the oral microbiome as noninvasive biomarkers for determining the diagnosis and prognosis of EC.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12967-025-06989-z. Introduction Esophageal cancer (EC) ranks as the seventh most common cancer and the sixth leading cause of cancer-related death worldwide, representing a significant public health threat [1]. The two primary pathological types of EC are esophageal squamous carcinoma and esophageal adenocarcinoma, with esophageal squamous carcinoma being more prevalent in China [2]. In recent years, the incidence of EC has been increasing. Due to the lack of early screening and diagnostic methods, most patients are diagnosed at intermediate to advanced stages, resulting in an overall 5-year survival rate of only 13–18%. In fact, 40% of patients present with distant metastases at the time of the initial diagnosis, which is a primary cause of treatment failure and death in EC patients. An early diagnosis enables timely intervention and is crucial for improving the clinical prognosis. However, the gold standard for diagnosing EC is endoscopy, an invasive and costly procedure that limits its widespread use [3, 4]. As a result, nonnvasive detection methods with high specificity and sensitivity are urgently needed. Furthermore, effective disease management through prognostic surveillance can significantly improve treatment outcomes. Currently, the clinical prognotic prediction relies heavily on doctors’ empirical judgment, which is based on various factors, such as the tumor location, TNM staging information, and histopathology, which have limited accuracy and consistency [5, 6]. Therefore, the development of a novel prediction method is essential for better disease management. The human microbiome plays critical roles in disease onset, disease progression, and response to treatment. Recent studies have highlighted an association between the oral microbiome and diseases, including local lesions (e.g., dental caries and periodontal disease), as well as systemic diseases, including cancer [7–9]. Consequently, the oral microbiome has been increasingly utilized to identify biomarkers and potential therapeutic targets for a range of diseases. In the past five years, research has advanced regarding role of the oral microbiome in the development of esophageal diseases, including precancerous lesions, esophageal squamous cell carcinoma, and esophageal adenocarcinoma [10]. Dysregulation of the oral microbiome may contribute to esophageal disease by modulating the tumor immune response, promoting inflammation, and influencing the abnormal expression of signaling molecules. However, challenges still remain, including the small size of queues and the lack of independent queue validation. Thus, the development of robust microbial biomarkers for clinical applications is essential. While the microbiome encompasses a wide range of microbes found in clinical samples, the processing and interpretation of such complex data remains a challenge. Machine learning (ML), a widely used artificial intelligence (AI) technique, has proven valuable in automating data analysis across numerous medical domains. It offers unique advantages in interpreting omics data, developing predictive models, identifying biomarkers, and stratifying patients for precision medicine [11, 12]. However, the application of ML in analyzing microbiome data and developing potential biomarkers for EC remains underexplored, underscoring the potential for further research in this area. The Central Plains region of China is a high-incidence area of EC. In this study, we collected tongue-coating swabs from 167 healthy controls (HCs), 153 EC patients, and 120 EC patients who received radiotherapy from Henan Province, China. We employed 16 S rRNA MiSeq sequencing to characterize the oral microbiome profiles of EC patients and alterations before and after radiotherapy. A noninvasive diagnostic model was constructed and validated using an ML-based analysis of the microbiome data. Additionally, EC patients who had undergone radiotherapy were followed for six months. Prognostic markers were identified and a prognostic model was constructed. Our results demonstrate the potential of ML to analyze microbiome data, facilitating noninvasive detection and advancing precision medicine for treating EC. Materials and methods Study design and flow diagram This study followed the principles of prospective specimen collection and retrospective blinded assessment design. A total of 153 pathologically confirmed EC patients and 167 HCs (cohort 1) were recruited from the Affiliated Cancer Hospital of Zhengzhou University from June 2022 to September 2023 (Fig. 1). All enrolled EC patients received radiotherapy after sample collection. Additionally, 120 EC patients provided samples after radiotherapy, allowing for the observation of oral microbiome dynamics before and after radiotherapy. The preradiotherapy and postradiotherapy samples were categorized into the nonradiotherapy (NRT) and radiotherapy (RT) groups, respectively (cohort 2). After applying strict inclusion and exclusion criteria, 440 samples were included for 16S rRNA MiSeq sequencing. Patients who received anticancer treatment before sampling were excluded. Clinical and demographic data were collected and analyzed. The Online Supplementary Methods describe inclusion, exclusion and detailed diagnostic criteria. The study adhered to all relevant ethical guidelines and was approved by the Ethics Committee of the Affiliated Cancer Hospital of Zhengzhou University (Institutional Review Board: 2022-KY-0108-003). Written informed consent was obtained from each participant. Tongue-coating sample collection and DNA extraction Before sample collection, all the participants rinsed their mouths twice with sterile mouthwash. A professional operator then scraped the anterior and posterior middle regions of the tongue using a pharyngeal swab. The samples were immediately placed in test tubes and transferred to a −80 °C freezer within 2 h. DNA extraction and PCR amplification were performed using standardized protocols by the same laboratory staff. The samples were suspended in 790 µL of sterile lysis buffer (4 M guanidine thiocyanate, 250 µL; 10% N-lauroyl sarcosine, 40 µL; 5% N-lauroyl sarcosine-0.1 M phosphate buffer [pH 8.0], 500 µL), vortexed vigorously and incubated at 70 °C for 1 h. At the end of the incubation, glass beads (0.1 mm, 500–750 µL) were added and mixed, and the beads were struck for 10 min (25 HZ/S) to sufficiently lyse the cell membrane and the nuclear membrane. Genomic DNA was extracted using the E.Z.N.A. ® Stool DNA Kit (Omega Bio-tek, Inc., GA) according to the manufacturer’s instructions and stored at −20 °C for further analysis. PCR amplification and sequencing The DNA extracted from each sample was used as a template to amplify the V3-V4 region of the 16S rRNA gene (primers 341F: 5’-CCTACGGGGNGGCWGCAG − 3’ and 805R: 5’-GACTACHVGGGTATCTAATCC-3’). PCR amplification was conducted using an EasyCycler 96 PCR system (Analytik Jena Corp., AG). Amplicons were purified with Hieff NGS® DNA Selection Beads (YeasenBiotechCo., Ltd., China). The products were mixed in the same proportions for sequencing by Shanghai Mobio Biomedical Technology Co. Ltd. in China using the Illumina MiSeq platform (Illumina Inc., USA). Bioinformatics and statistical analysis Amplicon sequence variants (ASVs) were identified using the DADA2 (Divisive Amplicon Denoising Algorithm 2) algorithm. The representative sequences for each ASV were annotated with the SILVA reference database (SSU138.2). Alpha diversity metrics (ACE estimator, Chao 1 estimator, Shannon‒Wiener diversity index and Simpson diversity index) were assessed by using Mothur v1.42.1. Differences between two groups were assessed using the nonparametric Mann‒Whitney U test, whereas multiple group comparisons were performed using the Kruskal‒Wallis test. The Bray‒Curtis, weighted and unweighted UniFrac dissimilarity were calculated with QIIME. Principal coordinate analysis (PCoA) plots and PERMANOVA which were used to test for the statistical significance of differences between the groups based on 10,000 permutations were generated in the R (version 3.6.0) package vegan 2.5-7. Linear discriminant analysis (LDA) effect size (LEfSe) was used to identify taxa with differential abundance among groups (LEfSe 1.1, https://github.com/SegataLab/lefse). A heatmap plot of the key ASVs identified by random forest models was generated using the ‘pheatmap’ package of the R program. PICRUSt2 v2.4.1 (https://github.com/picrust/picrust2/wiki) was used to predict functional abundances based on 16 S rRNA gene sequences. The raw sequencing data of the 16 S rRNA gene V3–V4 regions and accompanying information are available in Sequence Read Archive database under accession number PRJNA1214484. Diagnostic model ASV frequency profiles for each cohort were generated by mapping reads to the corresponding representative sequences. In the discovery cohort, significant differences between the EC and HC groups (P < 0.05) were identified using the Wilcoxon rank sum test. The ASVs identified from the discovery set were used to construct a diagnostic model based on the random forest algorithm. We ranked the ASVs according to their significance values (Mean decrease accuracy) from high to low, increased the number of ASVs in order, and calculated the error of constructing diagnostic models using different numbers of biomarkers to obtain the optimal combination of diagnostic markers. Fivefold cross-validation (CV) was conducted to obtain error curves. The mean of the CV errors was calculated from the data of the five curves to obtain a line graph plotted with the mean values. We chose the minimum CV error plus the standard deviation of the CV error from the mean curve as the critical value. We listed the set of all the ASV markers for which the CV error was less than the critical value, and selected the set with the least number of ASVs as the optimal set. Finally, the probability of disease (POD) indices of the discovery and validation sets were calculated from the identified optimal ASV sets. The POD index is the ratio of the number of randomly generated decision trees with predicted samples as EC patients to the number of predicted HCs. The model was then evaluated by constructing receiver operating characteristic (ROC) curves (R 3.3.0, pROC package) and calculating the area under the curve (AUC). Prognostic model A total of 132 patients from Cohort 3 with 6-month postoperative outcome data were randomly stratified to establish an EC prognostic model. Patients were categorized into a progression group (PG) (n = 26) and a nonprogression group (NPG) (n = 106) based on the presence of tumor recurrence, metastasis, and death during the follow-up period. A random survival forest (RSF) model with 1000 trees was trained, and salient features were selected based on replacement-based feature importance. The optimal model was established by retraining the RSF model using the 13 selected microbes, which showed a good predictive ability for the 6-month outcomes of EC patients. Statistical analysis Continuous variables are reported as the means (standard deviations) or medians (interquartile ranges), whereas categorical variables are presented as percentages. Student’s t-tests and Wilcoxon rank sum tests were used for normally and nonnormally distributed continuous variables, respectively. Chi-square tests or Fisher’s exact tests were applied to categorical variables. Statistical analyses were performed using SPSS V.20.0 for Windows (SPSS, Chicago, Illinois, USA), with P < 0.05 (two-tailed) considered statistically significant. Results Patients, data collection, and study design The overall workflow of this study and detailed participant recruitment information for each analysis are presented in Fig. 1. A total of 440 tongue-coating samples were prospectively collected from the Central Plains. After rigorous exclusion and diagnostic procedures, 440 samples were analyzed, including 157 EC patients, 167 HCs, and 120 EC patients after radiotherapy. The EC and HC samples were randomized into discovery and validation phases at a 7:3 ratio. In the discovery phase, we characterized the oral microbiome in 107 EC samples and 117 matched HC samples, identified key oral microbial markers, and constructed EC classifiers through a fivefold cross-validated random forest model. In the validation phase, we tested the diagnostic efficacy of the EC classifier based on oral microbiome from 50 EC samples. Additionally, we investigated the dynamics of the oral microbiome in nonradiotherapy (NRT) (120 samples) and radiotherapy (RT) patients (120 samples). Over a 6-month follow-up period, we used ML to analyze the oral microbiome data of 132 EC patients (cohort 3) to develop a prognostic model. Baseline characteristics of the study cohort The clinical characteristics of the EC patients and HCs in the discovery cohort are presented in Table 1 (Online Supplementary Table S1). Age- and gender-matched patients were compared between the EC and HC groups (P > 0.05). The neutrophil counts (NEUT) (P < 0.05) were increased, whereas lymphocyte counts (LYMPH) (P < 0.05), monocyte counts (MONO) (P < 0.05), hemoglobin (HGB) level (P < 0.001), and albumin (ALB) level (P < 0.001) were decreased in the EC group than in the HC group. Table 1Clinical characteristics of participants in the oral discovery cohortClinical indexesEC patients (n = 107)Healthy controls (n = 117)P-valueAge (years)65.56 ± 9.1664.80 ± 7.140.232Sex (female/male)33/7440/920.232Smoking50 (46.7%)NA··Drinking49 (45.8%)NA··ECOG score··029 (27.1%)NA178 (72.9%)NATumor Stage··I0NAII15 (14.02%)NAIII61 (57.01%)NAIV31 (28.97%)NAWBC6.03 (4.74, 7.12)5.80 (4.87, 6.60)0.428NEUT4.16 (2.99, 4.81)3.23 (2.63, 4.09)0.001LYMPH1.57 (1.22, 1.86)1.68 (1.47, 2.11)0.017MONO0.30 (0.23, 0.35)0.37 (0.28, 0.44)0.001HGB134 (123, 143)146.0 (134.5, 157.0)< 0.001PLT221 (176, 267)215.5 (189, 249.3)0.524ALB40.6 (37.7, 43.4)46.8 (44.7, 48.6)< 0.001 Table 1Clinical characteristics of participants in the oral discovery cohortClinical indexesEC patients (n = 107)Healthy controls (n = 117)P-valueAge (years)65.56 ± 9.1664.80 ± 7.140.232Sex (female/male)33/7440/920.232Smoking50 (46.7%)NA··Drinking49 (45.8%)NA··ECOG score··029 (27.1%)NA178 (72.9%)NATumor Stage··I0NAII15 (14.02%)NAIII61 (57.01%)NAIV31 (28.97%)NAWBC6.03 (4.74, 7.12)5.80 (4.87, 6.60)0.428NEUT4.16 (2.99, 4.81)3.23 (2.63, 4.09)0.001LYMPH1.57 (1.22, 1.86)1.68 (1.47, 2.11)0.017MONO0.30 (0.23, 0.35)0.37 (0.28, 0.44)0.001HGB134 (123, 143)146.0 (134.5, 157.0)< 0.001PLT221 (176, 267)215.5 (189, 249.3)0.524ALB40.6 (37.7, 43.4)46.8 (44.7, 48.6)< 0.001 Clinical characteristics of participants in the oral discovery cohort Characterization of the oral microbiome of EC patients The Venn diagram showed that 1578 out of 3065 ASVs were shared between the EC and HC groups, whereas1245 ASVs were unique to the EC group (Fig. 2a). The species accumulation curve and Shannon‒Wiener curve revealed nearly saturated ASV diversity and richness in each group (Fig. 2b and c). Alpha diversity, as measured by the Chao 1 index (P < 0.001), Ace index (P < 0.001) and Shannon’s index (P < 0.001), revealed greater oral microbial diversity and richness in the EC compared with HC (Figs. 2b, S1a and 2c, Online Supplementary Table S2). The beta diversity analysis using PCoA and NMDS plots calculated by unweighted UniFrac distance indicated the significant separation of microbial communities between the two groups, and Adonis (P = 0.001) and ANOSIM (P = 0.001) analyses showed that there were significant differences between the two groups (Fig. 2d and e, S1b, S1c and S1d). These findings suggest oral microbiome dysbiosis in EC. Further analysis of the composition and variation in EC patients and HCs revealed that Firmicutes, Bacteroidota, Proteobacteria, Actinobacteriota and Fusobacteriota were the five most abundant phyla in both groups, accounting for more than 95% of the sequences (Fig. S1e, Online Supplementary Table S3). The predominant genera included Neisseria, Prevotella, Veillonella, Streptococcus and Porphyromonas (Fig. 2f, Online Supplementary Table S4). At the phylum level, nine phyla, including Firmicutes, Fusobacteriota, Patescibacteria and Campilobacterota, were more abundant in the EC patients, whereas Bacteroidota and Proteobacteria were less abundant (Fig. S1f, Online Supplementary Table S5). At the genus level, 59 genera, including Veillonella, Streptococcus and Actinomyces, were enriched in EC, whereas 13 genera, including Porphyromonas, Rothia, and Haemophilus, were depleted (Fig. 2g, Online Supplementary Table S6). Linear discriminant analysis (LDA) effect size (LEfSe) analysis identified the most representative genera associated with EC based on LDA (LDA > 3.0, P < 0.05) (Fig. S1g, Online Supplementary Table S7). The heatmap revealed the relative abundances of top 50 ASVs for each sample in the EC patients and HCs, including Fusobacterium, Streptococcus, Lachnoanaerobaculum and Gemella (Fig. 2h, Online Supplementary Table S8). Fig. S1h showed the Spearman correlation between the top 50 ASVs in the EC patients and HCs. The random forest method identified differentially expressed ASVs in the EC patients and HCs, including Streptococcus, Gemella, Prevotella and Fusobacterium (Fig. S1i, Online Supplementary Table S9). These findings suggest a unique oral microbiome profile in the EC group. Machine learning-derived biomarker panel for diagnosing EC patients We next t used the obtained oral microbiome profiles to develop a novel diagnostic approach for EC. Through fivefold cross-validation in a random forest model, we selected six ASVs for distinguishing EC patients from HCs, including Alloprevotella, Prevotella, Campylobacter, Neisseria, Campylobacter and Neisseria (Figs. 3a, and 3b). Then, we established a 6-oral microbiome diagnostic model (6-OMD model). Using the six ASVs, we calculated the POD index of the discovery cohort (Online Supplementary Table S8). The POD index was significantly higher in EC patients than in HCs (Fig. 3c), with an AUC of 93.69% (95% CI: 89.2–98.71%, P < 0.0001) (Fig. 3d). Alloprevotellac and Prevotella were the two microbiome genera with the greatest contributions. These data suggest that oral microbial markers can specifically differentiate EC patients from HCs. The diagnostic efficacy of the microbial markers was validated in the validation cohort of 50 EC patients, with the POD index showing a significantly higher value for EC patients and an AUC of 95.18% (95% CI: 92.47–97.89%, P < 0.0001) (Fig. 3e and f, Online Supplementary Table S11). These results suggest that oral microbiome-based markers can serve as a noninvasive tool for diagnosing EC. Dynamic changes in the oral microbiome of EC patients before and after radiotherapy We examined and compared the characteristics of the oral microbiome of 120 NRT group and matched 120 matched in the RT group. Venn diagram revealed that of the 2938 ASVs, 2026 were common to the RT and NRT groups, whereas 229 and 683 ASVs were unique to the RT and NRT groups, respectively (Fig. 4a). Compared with those in the NRT group, the oral microbial diversity and richness in the RT groups were significantly reduced (P < 0.05) (Fig. 4b and c, Online Supplementary Table S12). The PCoA and NMDS analyses indicated significant differences in the distribution of the oral microbiome between the RT and NRT groups (P = 0.001, ANOSIM) (Fig. 4d and e, S2a, Online Supplementary Table S13). The mean composition and relative abundance of the oral microbiome at the phylum and genus levels are presented in Figs. S2b and 4f (Online Supplementary Tables S14 and S15). The predominant microbe in the RT group was Streptococcus, while the predominant microbe in the NRT group was Neisseria. Further analysis of differences at the genus and phylum levels showed that 45 genera differed significantly between the two groups (all P < 0.05), with 41 genera, including Fusobacterium and Porphyromonas, decreasing in abundance after radiotherapy, (Figs. 4g and S2c, Online Supplementary Tables S16 and S17). In contrast, the abundances of four genera, Streptococcus, Actinomyces, Corynebacterium and Pediococcus, increased. The differential microbial analysis confirmed that the microbial profile of the RT group differed significantly from that of the NRT group (Fig. 4h). LEfSe identified the most representative genera of both groups based on LDA score (LDA >3.0, P < 0.05) (Fig. S2d, Online Supplementary Table S18). Oral microbial characterization of the PG and NPG Clinical information from 132 EC patients (cohort 3) was collected over a follow-up period of 6 months. An UpSet plot showed that 1271 out of 2753 ASVs were common to both groups, whereas 233 ASVs were unique to the PG (Fig. 5a). The oral microbial diversity and abundance were similar between the PG and the NPG (P >0.05) (Fig. 5b, Online Supplementary Table S19). Although the Shannon index did not differ significantly, it decreased in the PG. The beta diversity analysis using PCoA and NMDS did not distinguish between different EC patients (Fig. 5c and d). These findings suggest that the microbial characteristics associated with oral diversity are similar between the PG and NPG. Figures 4e and S2e illustrate the mean composition and relative abundance of the oral microbiome at the genus and phylum levels (Online Supplementary Tables S20 and S21). Despite similar taxonomic characteristics, specific differential bacteria were identified. Compared with those in the NPG, the abundances of four genera, Fusobacterium, Eubacterium_nodatum_group, Lachnospiraceae_uncultured, and Filifactor, were reduced in the PG (all P < 0.05) (Fig. 4f, Online Supplementary Table S22). The random forest analysis identified six key differentially abundant ASVs, including Fusobacterium and Gemella (Fig. 4g, Online Supplementary Table S23). We further analyzed the correlation between the oral microbiome and EC clinical data by Spearman correlation analysis. As shown in Figs. 5h and 25 ASVs were correlated with nine clinical indicators including WBC, NEUT, PLT, HGB levels, tumor stage, smoking, LYM, MONO and ALB levels (Fig. 4h, Online Supplementary Table S24). Overall, these results suggested that while the abundance and composition of the oral microbiome in the PG were generally consistent with those in the NPG, some specific bacterial changes were more pronounced in the PG. A prognostic model based on the oral microbiome accurately predicts 6-month outcomes in EC patients Since an accurate prognosis enables precise interventions and benefits patients’ clinical outcomes, we developed an ML-based prognostic model. The diagnostic efficacy of the model was assessed by constructing ROC curves, based on the differences in ASVs between the PG and the NPG in the preradiotherapy samples. We established a 13-oral microbiome prognostic model (13-OMP model) using the RSF approach. The POD index was significantly higher in the NPG than in the PG (Fig. 6b, Online Supplementary Table S25), with an AUC of 99.29% (95% CI: 98.41–100%) (Fig. 6a). Considering the impacts of clinical indicators on the prognostic prediction, we further modeled prognosis based on clinical features and differentially abundant ASVs. However, integrating the clinical features did not improve the model’s predictive performance, which achieved an AUC of 97.68% (95% CI: 95.57–99.8%) compared with that of the 13-OMP model (Fig. 6e). This result highlights the robust predictive ability of our model. Predictive functions of the critical oral microbiome correlated with EC progression The enriched pathways of the oral microbiome that may influence EC progression were identified in the KEGG database. Compared with those in HCs, 27 functional modules (including glyoxylate and dicarboxylate metabolism, pentose and glucuronate interconversions, and pentose phosphate pathway) were significantly increased in EC patients, whereas 22 functions, such as lipoic acid metabolism, glutathione metabolism and the citrate cycle, were significantly decreased (LDA >3, P < 0.05) (Fig. 7a, Online Supplementary Table S26). In the RT group, 27 functional modules, including amino sugar and nucleotide sugar metabolism, glycolysis and gluconeogenesis, and the pentose phosphate pathway, were significantly increased compared with those in the NRT group, whereas 18 functional modules, such as riboflavin metabolism, tyrosine metabolism, and biosynthesis of unsaturated fatty acids, were significantly reduced (LDA >3, P < 0.05) (Fig. 7b, Online Supplementary Table S27). The MetaCyc database identified enriched pathways in the oral microbiome that may affect the EC prognosis. The results showed six pathways were enriched in the PG, such as glycolysis I, glycolysis II, and amino acid biosynthesis, while five pathways, including pyruvate fermentation, L glutamate degradation, and L lysine fermentation, were depleted (LDA >3, P < 0.05) (Fig. 7c, Online Supplementary Table S28). These findings suggest the dysregulation of functional patterns in EC progression. Discussion In this study, we employed ML algorithms to analyze microbiome data, aiming to further develop two biomarker panels: the 6-OMD model and the 13-OMP model, both of which demonstrated high accuracy. We utilized a clinical cohort from the Central Plains to investigate oral microbial alterations between EC patients and HCs, and identified key bacteria and microbial functions of potential diagnostic value. Additionally, we examined changes in the oral microbiome of EC patients before and after radiotherapy through follow-up. Ultimately, oral microbes of potential prognostic value were identified based on a 6-month follow-up. Our findings highlight the unique advantages of applying an ML-based microbiome for the early detection of EC and precision medicine, which has significant clinical translational potential. Any change in the composition or function (in terms of abundance and diversity) of the resident commensal microbiome may lead to inflammation and metabolic abnormalities, potentially resulting in variety of diseases [13]. Studies on oral disease and cancer provide evidence that oral health issues (tooth loss, poor oral hygiene and possibly periodontal disease) are associated with the risk of EC [14]. Oral bacterial overgrowth may play a critical role in the development and establishment of EC, potentially because of an increase in multiple pathogenic or opportunistic pathogens. For the first time, we reported oral microbial dysbiosis in a large sample of EC patients. The oral bacterial abundance and diversity indices were significantly increased in the EC patients, indicating increased bacterial colonization. Variations in diet and environment may have contributed to these differing results across studies. The relative abundances of Veillonella, Streptococcus and Actinomyces increased in EC patients, whereas the relative abundances of Porphyromonas, Rothia and Haemophilus decreased. A prior study has highlighted the enrichment of Streptococcus in EC tissues, which enhanced the efficacy of anti-PD-1 immunotherapy [15]. Another study noted an increase in Actinomyces, and a decrease in Streptococcus in patients with a postsurgical recurrence of EC compared with those without recurrence [16]. These microbes may be preferred targets for interventional therapy in EC patients. We also found that the predicted microbial functions of the oral microbiome differed between the two groups, with the glyoxalate and dicarboxylate metabolic pathways enriched in EC patients. A metabolomic approach has linked this pathway to chemoradiotherapy sensitivity, providing new therapeutic ideas [17]. Specific oral microbiomes can be used as diagnostic markers, which have been investigated in a number of diseases [18, 19], although similar models for EC remain limited. Our study developed and validated an EC classifier using six ASVs of the oral microbiome, achieving AUCs of 93.69% and 95.18% in discovery and validation groups, respectively. This study is the first to successfully identify oral microbiome-based EC classifiers and achieve independent validation, highlighting its potential as a noninvasive diagnostic tool. The microbiome changes as the disease progress. One study revealed that the oral microbiome partially normalized after periodontitis treatment [20]. Our previous research showed that the oral microbiome was associated with recovery from COVID-19 [21]. In this study, we evaluated changes in the oral microbiome of EC patients before and after radiotherapy, and found significant differences in microbial diversity and richness postradiotherapy. The prevalence of opportunistic pathogens, including Fusobacterium and Porphyromonas, was significantly decreased in the RT group, which was associated with progression and poor prognosis of EC [22, 23]. LEfSe indicated a significant increase in the functional module of the pentose phosphate pathway in the RT group. Activation of the pentose phosphate pathway increases radiotherapy sensitivity [24]. These findings suggest that radiotherapy may promote EC recovery by altering the diversity and abundance of oral bacteria. In addition to their diagnostic capabilities, we identified new microbial prognostic biomarkers. Prognostic models for EC based on the oral microbiome remain limited. Through six months follow-up, we developed a prognostic model based on 13 selected ASVs of the oral microbiome. Considering the influence of clinical indicators on the prognostic prediction, we further developed prognostic models based on clinical features and differentially abundant ASVs and found that they did not improve the predictive performance of the models. This model enables the early identification of EC patients at risk of progression postradiotherapy, allowing for personalized clinical decisions. Future validation and optimization with larger cohorts and advanced ML algorithms are needed. Moreover, we identified six pathways enriched in the PG, such as glycolysis I, and glycolysis II. Enhanced glycolysis pathways can promote EC and attenuate radiosensitivity [25, 26]. This finding suggested that the progression may be due to enhanced glycolysis reducing the efficacy of radiation therapy. A study revealed that Akkermansia muciniphila and its metabolite enhanced oxaliplatin efficacy in gastric cancer patients by modulating glycolysis [27]. These findings provide a new perspective on regulating glycolysis through probiotic or prebiotic intervention. Although our study illustrated the role of the oral microbiome in EC, it has several limitations. First, this study was conducted at a single-center. The universality of the findings needs further validation by performing studies in other regions. Second, the follow-up period for patients was relatively short. Future studies should extend the follow-up period to explore the relationship between the oral microbiome and prognosis. Third, this study did not clarify the causal relationship between the oral microbiome and esophageal cancer. Additional experiments involving microbial-related factors are needed to validate our findings. Conclusion This study represents the first cross-sectional and longitudinal analysis of the oral microbiome in EC patients. We analyzed characteristic alterations in the oral microbiome in EC patients from a large cohort, identified key bacteria and predicted specific microbial functions associated with EC. These results suggest that the dysregulation of the oral microbiome may contribute to the development of EC. We developed two ML-based models for EC diagnosis and prognosis. In addition, we characterized changes in the oral microbiome of EC patients before and after radiotherapy, suggesting that the oral microbiome may be associated with disease recovery. Our work advances the understanding of the relationship between the oral microbiome and EC, highlighting the use of oral microbiome as a therapeutic target for future EC treatments. Fig. 1Overview of the study design. A total of 320 individuals were included in the study, and their tongue-coating samples were subjected to a targeted metabolomics analysis and 16 S rRNA MiSeq sequencing. The microbiome profiles of esophageal cancer (GC) patients and healthy controls (HCs) in Cohort 1 (n = 320) were compared to characterize the oral microbiome of EC patients. Using the oral microbiome data from cohort 1 and machine learning techniques, a diagnostic model for EC (6-OMD model) was created and validated. Among the EC patients, 120 patients had samples collected after radiotherapy; these samples were used to observe the dynamics of oral microbiome before and after radiotherapy (cohort 2, n = 240). Oral microbiome data from cohort 3 (n = 132) patients and their clinical features were analyzed using a machine learning algorithm to develop a prognostic model (13-OMP model). The different colored triangles in the figure represent various participant groups used for model construction, validation, and comparison processes. EC, esophageal cancer; HC, healthy control; RT, radiotherapy; NRT, nonradiotherapy; rRNA, ribosomal RNA Overview of the study design. A total of 320 individuals were included in the study, and their tongue-coating samples were subjected to a targeted metabolomics analysis and 16 S rRNA MiSeq sequencing. The microbiome profiles of esophageal cancer (GC) patients and healthy controls (HCs) in Cohort 1 (n = 320) were compared to characterize the oral microbiome of EC patients. Using the oral microbiome data from cohort 1 and machine learning techniques, a diagnostic model for EC (6-OMD model) was created and validated. Among the EC patients, 120 patients had samples collected after radiotherapy; these samples were used to observe the dynamics of oral microbiome before and after radiotherapy (cohort 2, n = 240). Oral microbiome data from cohort 3 (n = 132) patients and their clinical features were analyzed using a machine learning algorithm to develop a prognostic model (13-OMP model). The different colored triangles in the figure represent various participant groups used for model construction, validation, and comparison processes. EC, esophageal cancer; HC, healthy control; RT, radiotherapy; NRT, nonradiotherapy; rRNA, ribosomal RNA Fig. 2Oral microbial dysbiosis in EC patients. (a) Venn diagram showing that 1578 of the 3065 ASVs were common to the EC and HC groups, and 1245 ASVs were unique to the EC group. (b) Rarefaction analysis between the number of samples and the number of ASVs. As the number of samples increased, the number of ASVs tended to saturate. The number of ASVs increased in EC patients compared with HCs. Compared with that of HCs, the oral microbial richness of EC patients was significantly increased according to the Chao 1 index. (c) The Shannon‒Wiener curves of the samples showed that the Shannon diversity approached saturation as the number of sequences increased. Compared with that of HCs, the oral microbial diversity of the EC patients increased, as estimated by the Shannon index. PCoA (d) and NMDS (e) analyses revealed that the taxonomic compositions of the oral microbes in both groups were obviously different. (f) Average composition and relative abundance of the two bacterial communities at the genus level. (g) At the genus level, 59 microbes were enriched in the EC group, and 13 microbial features were depleted. (h) Heatmap of the relative abundances of top 50 ASVs for each sample in the EC patients and HCs. EC, esophageal cancer; HC, healthy control; ASVs, amplicon sequence variants; PCoA, principal coordinate analysis; NMDS, nonmetric multidimensional scaling. *P < 0.05, ** P < 0.01, and *** P < 0.001 Oral microbial dysbiosis in EC patients. (a) Venn diagram showing that 1578 of the 3065 ASVs were common to the EC and HC groups, and 1245 ASVs were unique to the EC group. (b) Rarefaction analysis between the number of samples and the number of ASVs. As the number of samples increased, the number of ASVs tended to saturate. The number of ASVs increased in EC patients compared with HCs. Compared with that of HCs, the oral microbial richness of EC patients was significantly increased according to the Chao 1 index. (c) The Shannon‒Wiener curves of the samples showed that the Shannon diversity approached saturation as the number of sequences increased. Compared with that of HCs, the oral microbial diversity of the EC patients increased, as estimated by the Shannon index. PCoA (d) and NMDS (e) analyses revealed that the taxonomic compositions of the oral microbes in both groups were obviously different. (f) Average composition and relative abundance of the two bacterial communities at the genus level. (g) At the genus level, 59 microbes were enriched in the EC group, and 13 microbial features were depleted. (h) Heatmap of the relative abundances of top 50 ASVs for each sample in the EC patients and HCs. EC, esophageal cancer; HC, healthy control; ASVs, amplicon sequence variants; PCoA, principal coordinate analysis; NMDS, nonmetric multidimensional scaling. *P < 0.05, ** P < 0.01, and *** P < 0.001 Fig. 3Identification of an oral microbial classifier for EC. (a) Six microbial markers were selected as the best set of markers by random forest model. (b) Distribution map of the importance of the selected microbial markers in the model. The POD values were significantly increased in the EC patients compared with HCs, and achieved great efficacy in the discovery cohort (c and d) and the validation cohort (e and f). EC, esophageal cancer; HC, healthy control; POD, probability of disease; AUC, area under the curve. Centerline, median; box limits, upper and lower quartiles; circle or square symbol, mean; error bars, 95% CI Identification of an oral microbial classifier for EC. (a) Six microbial markers were selected as the best set of markers by random forest model. (b) Distribution map of the importance of the selected microbial markers in the model. The POD values were significantly increased in the EC patients compared with HCs, and achieved great efficacy in the discovery cohort (c and d) and the validation cohort (e and f). EC, esophageal cancer; HC, healthy control; POD, probability of disease; AUC, area under the curve. Centerline, median; box limits, upper and lower quartiles; circle or square symbol, mean; error bars, 95% CI Fig. 4Alterations in the oral microbiome of EC patients before and after radiotherapy. (a) Venn diagram showing that of the 2938 ASVs, 2026 were common to the RT and NRT groups, whereas only 229 and 683 ASVs were unique to the RT and NRT groups, respectively. (b) Rarefaction analysis between the number of samples and the number of ASVs. As the number of samples increased, the number of ASVs tended to saturate. As estimated by the Chao index, the diversity and abundance of the oral microbiome in the RT group were significantly reduced compared with the NRT group. (c) Shannon-Wiener curves of the samples showing the saturation of Shannon diversity as the number of sequences increased. The diversity and richness of the oral microbiome in the RT group were lower than in the NRT group. PCoA (d) and NMDS (e) revealed that the distribution of the oral microbiome in the RT group was significantly different from that in the NRT group. (f) Average compositions and relative abundances of the bacterial community in the RT and NRT groups at the genus level. (g) Compared with the NRT group, 41 genera were depleted in the RT group. (h) Heatmap of the relative abundances of differentially abundant ASVs for each sample in the RT and NRT groups. RT, radiotherapy; NRT, nonradiotherapy; ASVs, amplicon sequence variants; PCoA, principal coordinate analysis; NMDS, nonmetric multidimensional scaling. *P < 0.05, ** P < 0.01, and *** P < 0.001 Alterations in the oral microbiome of EC patients before and after radiotherapy. (a) Venn diagram showing that of the 2938 ASVs, 2026 were common to the RT and NRT groups, whereas only 229 and 683 ASVs were unique to the RT and NRT groups, respectively. (b) Rarefaction analysis between the number of samples and the number of ASVs. As the number of samples increased, the number of ASVs tended to saturate. As estimated by the Chao index, the diversity and abundance of the oral microbiome in the RT group were significantly reduced compared with the NRT group. (c) Shannon-Wiener curves of the samples showing the saturation of Shannon diversity as the number of sequences increased. The diversity and richness of the oral microbiome in the RT group were lower than in the NRT group. PCoA (d) and NMDS (e) revealed that the distribution of the oral microbiome in the RT group was significantly different from that in the NRT group. (f) Average compositions and relative abundances of the bacterial community in the RT and NRT groups at the genus level. (g) Compared with the NRT group, 41 genera were depleted in the RT group. (h) Heatmap of the relative abundances of differentially abundant ASVs for each sample in the RT and NRT groups. RT, radiotherapy; NRT, nonradiotherapy; ASVs, amplicon sequence variants; PCoA, principal coordinate analysis; NMDS, nonmetric multidimensional scaling. *P < 0.05, ** P < 0.01, and *** P < 0.001 Fig. 5Oral microbial characterization of the PG and NPG. (a) An UpSet plot showing that 1271 out of 2753 ASVs were common to both groups, whereas 233 ASVs were unique to the PG. (b) Shannon‒Wiener curves of the samples showed the saturation of Shannon diversity as the number of sequences increased. The oral microbial diversity and abundance in the PG and NPG were similar. The PCoA (c) and NMDS (d) analyses were unable to distinguish between different EC patients. (e) Average compositions and relative abundance of the bacterial community in the PG and NPG at the genus level. (f) Compared with those in the NPG, 4 genera were depleted in the PG. (g) Heatmap of six key differentially abundant ASVs identified by the random forest analysis in the PG and NPG. (h) Correlations between the oral microbiome and clinical indicators. PG, progression group; NPG, nonprogression group; ASVs, amplicon sequence variants; PCoA, principal coordinate analysis; NMDS, nonmetric multidimensional scaling; WBC, white blood cell; NEUT, neutrophil; LYMPH, lymphocyte; Hb, hemoglobin; PLT, platelet; Alb, albumin; Mono, monocyte. *P < 0.05, ** P < 0.01, and *** P < 0.001 Oral microbial characterization of the PG and NPG. (a) An UpSet plot showing that 1271 out of 2753 ASVs were common to both groups, whereas 233 ASVs were unique to the PG. (b) Shannon‒Wiener curves of the samples showed the saturation of Shannon diversity as the number of sequences increased. The oral microbial diversity and abundance in the PG and NPG were similar. The PCoA (c) and NMDS (d) analyses were unable to distinguish between different EC patients. (e) Average compositions and relative abundance of the bacterial community in the PG and NPG at the genus level. (f) Compared with those in the NPG, 4 genera were depleted in the PG. (g) Heatmap of six key differentially abundant ASVs identified by the random forest analysis in the PG and NPG. (h) Correlations between the oral microbiome and clinical indicators. PG, progression group; NPG, nonprogression group; ASVs, amplicon sequence variants; PCoA, principal coordinate analysis; NMDS, nonmetric multidimensional scaling; WBC, white blood cell; NEUT, neutrophil; LYMPH, lymphocyte; Hb, hemoglobin; PLT, platelet; Alb, albumin; Mono, monocyte. *P < 0.05, ** P < 0.01, and *** P < 0.001 Fig. 6A prognostic model based on the oral microbiome accurately predicts 6-month outcomes of patients with EC. (a and b) POD values were significantly decreased in the PG compared with the NPG, and achieve greater efficacy in the discovery cohort. (c) Variable importance of the 13 selected microbial markers in the model. (d and e) The POD values were significantly decreased in the PG compared with the NPG, and achieved great efficacy in the discovery cohort. (f) Variable importance based on 6 selected markers in the model according to the clinical features and differentially abundant ASVs A prognostic model based on the oral microbiome accurately predicts 6-month outcomes of patients with EC. (a and b) POD values were significantly decreased in the PG compared with the NPG, and achieve greater efficacy in the discovery cohort. (c) Variable importance of the 13 selected microbial markers in the model. (d and e) The POD values were significantly decreased in the PG compared with the NPG, and achieved great efficacy in the discovery cohort. (f) Variable importance based on 6 selected markers in the model according to the clinical features and differentially abundant ASVs Fig. 7The predicted functions of critical oral microbes are correlated with EC. (a) Compared with those in HCs, 27 functional modules were significantly increased in EC patients, whereas 22 functions were significantly decreased (LDA >3, P < 0.05). (b) Compared with the NRT group, 27 functional modules were significantly increased in the RT group, whereas 18 functional modules were significantly reduced (LDA >3, p < 0.05). (c) According to the MetaCyc database, six pathways were enriched in the PG whereas five pathways were depleted (LDA >3, P < 0.05). EC, esophageal cancer; HC, healthy control; RT, radiotherapy; NRT, nonradiotherapy; PG, progression group; NPG, nonprogression group; LDA, linear discriminant analysis The predicted functions of critical oral microbes are correlated with EC. (a) Compared with those in HCs, 27 functional modules were significantly increased in EC patients, whereas 22 functions were significantly decreased (LDA >3, P < 0.05). (b) Compared with the NRT group, 27 functional modules were significantly increased in the RT group, whereas 18 functional modules were significantly reduced (LDA >3, p < 0.05). (c) According to the MetaCyc database, six pathways were enriched in the PG whereas five pathways were depleted (LDA >3, P < 0.05). EC, esophageal cancer; HC, healthy control; RT, radiotherapy; NRT, nonradiotherapy; PG, progression group; NPG, nonprogression group; LDA, linear discriminant analysis Supplementary Information Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2 Supplementary Table S1-S28 Supplementary Table S1-S28