Oral squamous cell carcinoma (OSCC), the most common malignancy of the oral cavity, continues to claim a heavy toll worldwide. According to GLOBOCAN 2022 estimates cited in a new review, lip and oral cavity cancers accounted for roughly 389,485 new cases and 188,230 deaths globally, and the five-year survival rate has remained stuck at approximately 50 percent, largely because about 60 percent of patients are diagnosed at advanced stages. A team of researchers from Peking University School and Hospital of Stomatology and collaborating institutions has now published a systematic synthesis of how the oral microbiome—the second-largest microbial reservoir in the human body, hosting more than 700 species—could change that picture by powering non-invasive, saliva-based risk assessment models.
The review, published in Frontiers in Microbiology, organizes a rapidly growing but fragmented literature into four distinct categories of predictive models that the authors insist should not be conflated. The first comprises detection or discrimination models that separate prevalent OSCC from healthy or non-malignant states. The second covers models that predict malignant transformation from oral potentially malignant disorders (OPMDs), such as leukoplakia and lichen planus, into invasive carcinoma. The third predicts recurrence or lymph node metastasis after treatment, and the fourth addresses prognosis and survival. Each category, the authors argue, differs fundamentally in biological basis, clinical context, study design, and validation requirements—a distinction that much of the existing literature has blurred.
The biological foundation for these models rests on consistent observations of microbial dysbiosis in OSCC patients. Cross-regional meta-analyses have identified stable signatures, with enrichment of genera including Fusobacterium, Porphyromonas, Prevotella, Peptostreptococcus, and Parvimonas, alongside depletion of commensal taxa such as Neisseria and Corynebacterium. Beta diversity analyses clearly distinguish OSCC from healthy states across saliva, swab, and tissue samples, even though findings on alpha diversity remain inconsistent, with some studies reporting increased richness due to niches created by tumor necrosis and others reporting decreased diversity from pathogen overgrowth. Fusobacterium nucleatum enrichment emerges as the most consistently reported finding, documented in East Asian, European, and North American cohorts, though the magnitude of enrichment and the specific species identified vary across studies.
Mechanistically, the review details how these microbes may promote tumorigenesis through several converging routes. Chronic inflammation serves as a critical bridge: P. gingivalis and F. nucleatum persistently activate host immune responses via Toll-like receptors, triggering NF-κB and MAPK signaling and the release of pro-inflammatory cytokines such as IL-1β, IL-6, and TNF-α. The two pathogens display active synergy, with F. nucleatum acting as a structural bridge facilitating P. gingivalis invasion, which in turn upregulates IL-6/STAT3 signaling and matrix metalloproteinases. Direct physical interactions also perturb oncogenic pathways: the FadA adhesin of F. nucleatum binds epithelial E-cadherin, activating β-catenin signaling and upregulating proto-oncogenes such as c-Myc and Cyclin D1, while the microbe’s Fap2 protein binds the inhibitory receptor TIGIT on natural killer cells, facilitating immune escape. Metabolically, bacteria and especially Candida albicans, which exhibits exceptionally high alcohol dehydrogenase activity, convert ethanol into acetaldehyde, a potent carcinogen that induces DNA adducts.
On the modeling side, the field has evolved from rudimentary univariate statistics toward machine learning, deep learning, and multimodal data fusion. Traditional approaches—logistic regression and Cox proportional hazards models—remain valued for their interpretability, but random forest, support vector machines, and XGBoost have become cornerstone tools for handling the high-dimensional, compositional nature of microbiome data and modeling complex non-linear relationships. Deep learning approaches, including convolutional neural networks that convert microbial classification trees into image-like structures and graph neural networks that capture relationships between data points, have shown promise in adjacent fields, but the review notes that OSCC-specific deep learning applications remain limited and largely conceptual—a research gap rather than an established methodology.
Representative performance figures illustrate why the approach has generated excitement. A cross-cohort diagnostic model using XGBoost and random forest on 1,255 saliva and swab samples achieved area under the receiver operating characteristic curve (AUROC) values of 0.88 to 0.94, with sensitivity of 86 percent and specificity of 89 percent, and retained an AUROC of 0.82 under leave-one-cohort-out validation. A metatranscriptomic random forest model on 160 subjects reached an AUROC of 0.92 for distinguishing OSCC from healthy states and 0.85 for distinguishing OSCC from OPMD. For recurrence prediction, an XGBoost model on 112 post-operative patients achieved an AUROC of 0.87, outperforming traditional clinicopathological parameters such as surgical margin and lymph node status, which achieved 0.71. A random forest model on 88 patients predicted occult lymph node metastasis with an AUROC of 0.84 using enrichment of Prevotella, Bifidobacterium, and Finegoldia.
Yet the review is emphatic that these numbers describe retrospective discrimination, not validated clinical utility. As of 2026, the overwhelming majority of OSCC microbiome models remain at the internally validated stage, tested within the same cohort that generated the training data, with limited external validation and virtually no prospective validation. Sample sizes remain small, with most studies below 200 participants; external validation beyond the development cohort is rare; clinical comparators are inconsistently reported; and calibration and decision-curve analyses are almost entirely absent. Dedicated malignant transformation prediction models are particularly scarce: most studies compare established OSCC with OPMD rather than prospectively following OPMD patients to identify transformation risk factors, leaving the population most relevant for preventive screening underserved.
The authors also flag a subtler interpretive hazard: reverse causation. Established OSCC lesions modify local oxygen tension, inflammation, nutritional availability, bleeding, and tissue architecture, thereby reshaping the microbiome. A microbial signature that distinguishes an established tumor from healthy mucosa may therefore represent a consequence of disease rather than a cause or early predictor—a distinction with profound implications for screening models that aim to detect disease before clinical manifestation. The prognostic literature adds further complexity. While high intratumoral F. nucleatum loads have been associated with better outcomes in mixed-stage cohorts, including lower recurrence rates and prolonged disease-specific survival (hazard ratio 0.42 in one study of 145 patients), the organism has been identified as a poor prognostic factor in early-stage HPV-negative tongue cancers, where it carried a hazard ratio of 2.15 for worse survival. The review frames this paradox as an interkingdom phenomenon specific to HPV-negative tongue cancers driven by bacterial-inflammatory pathways, underscoring the need for stratified analyses.
Methodological challenges compound these biological ones. Most 16S rRNA sequencing studies report relative abundance, which is subject to the closed-sum constraint: an increase in one taxon’s proportional representation necessarily decreases that of others even if absolute counts remain stable, potentially generating spurious associations. The review calls for quantitative PCR, flow cytometry, or spike-in controls to measure absolute microbial load. Confounding by geography, ethnicity, diet, tobacco, alcohol, periodontal health, and antibiotic exposure is often inadequately controlled, and prior antifungal and antiviral exposures are entirely absent from current confounder discussions despite their documented effects on microbial equilibrium. Information leakage—performing feature selection or normalization on the full dataset before cross-validation—artificially inflates performance estimates, and most studies do not fully report their prevention strategies. The near-exclusive reliance on 16S rRNA sequencing also means current models are bacteriome-centric, blind to fungi, viruses, and archaea, even though Candida albicans and high-risk HPV strains play active roles in OSCC pathogenesis.
Infrastructure gaps round out the obstacles. Authoritative repositories such as the expanded Human Oral Microbiome Database (eHOMD), the Human Microbiome Project, the Oral Microbiome Bank of China, and the Cancer Microbiome Atlas provide valuable reference data, but no centralized, globally standardized database dedicated exclusively to OSCC—combining longitudinal saliva and swab sampling with standardized clinical, radiomic, and pathological metadata—currently exists. The review’s roadmap for clinical translation prioritizes multicenter standardized data collaboration, prospective validation studies described as the single most important priority, multimodal integration of microbiome data with metabolomics, clinical features, and radiomics, and explainable artificial intelligence techniques such as LIME and SHAP to make model decisions transparent to clinicians. Microbiome-modulating interventions, including bacteriophage therapy and anti-Candida strategies, remain experimental, with none validated for OSCC prevention or treatment. For now, the message is measured: oral microbiome models represent a promising frontier for non-invasive screening and risk stratification, but moving from statistical association to clinically actionable tools will demand rigorous methodological standards, larger annotated cohorts, and evidence gathered prospectively in real clinical settings.
Subject of Research: Risk assessment models for oral squamous cell carcinoma based on oral microbiome databases
Article Title: Current research status of risk assessment models for oral squamous cell carcinoma based on microbiome databases
Article References: Li, G.-L., Zhang, Z.-C., Lyu, X.-M., Li, M.-X., Lin, P.-Y., Yu, H.-S., Zhu, Y.-Y., Zhang, G.-H., Duan, C.-G., Cheung, Y., Li, W., Zhang, Z.-J., Xu, T.-S., Chen, F., & Huang, M.-W. (2026). Current research status of risk assessment models for oral squamous cell carcinoma based on microbiome databases. Frontiers in Microbiology, 17, 1894946. https://doi.org/10.3389/fmicb.2026.1894946
Image Credits: AI Generated
DOI: 10.3389/fmicb.2026.1894946
Keywords: oral squamous cell carcinoma, oral microbiome, dysbiosis, Fusobacterium nucleatum, machine learning, 16S rRNA sequencing, biomarkers, early detection, malignant transformation, recurrence prediction, prognosis, multimodal integration
News Source: Nathaniel Bowman. (October 26, 2026). Oral Microbiome Models Show Promise for Detecting Mouth Cancer, but Clinical Validation Lags. Scienmag.
