Artificial intelligence has spread rapidly through oral microbiome and dental genomics research, but the evidence is dispersed across imaging, sequencing and multi-omics studies of uneven methodological quality. This narrative review organizes the literature as an evidence matrix and synthesizes the applications, the quality of the evidence and its consistency across three domains: machine learning applied to the oral microbiome, machine learning applied to dental genomics, and the integration of both into clinical prediction. Sources were searched in Scopus, PubMed and Crossref, and studies were extracted into a matrix with study, design, population, finding and quality. Of the fourteen studies that applied machine learning to the oral microbiome, supervised classification and regression predominated, with modest to good discrimination for age, malodour, substance use and disease status. Of the seven studies in the dental-genomics domain, genome-wide and transcriptome-wide approaches identified candidate genes and loci, but shared prioritization between systemic and periodontal traits remained limited. Of the eighteen studies in the diagnostic-imaging domain, convolutional networks achieved the highest reported accuracy for caries and oral cancer, although few studies used public datasets or external validation. The evidence supports feasibility rather than clinical readiness; methodological heterogeneity, small samples and scarce external validation were the most frequent limitations.
La inteligencia artificial se ha extendido con rapidez por la investigación del microbioma oral y la genómica dental, pero la evidencia está dispersa entre estudios de imagen, secuenciación y multi-ómica de calidad metodológica desigual. Esta revisión narrativa organiza la literatura como una matriz de evidencia y sintetiza las aplicaciones, la calidad de la evidencia y su consistencia en tres dominios: el aprendizaje automático aplicado al microbioma oral, el aplicado a la genómica dental y la integración de ambos en la predicción clínica. Las fuentes se buscaron en Scopus, PubMed y Crossref, y los estudios se extrajeron a una matriz con estudio, diseño, población, hallazgo y calidad. De los catorce estudios que aplicaron aprendizaje automático al microbioma oral predominaron la clasificación y la regresión supervisadas, con discriminación modesta a buena para la edad, el mal olor, el consumo de sustancias y el estado de enfermedad. De los siete estudios del dominio de genómica dental, los enfoques de genoma completo y de transcriptoma identificaron genes y loci candidatos, aunque la priorización compartida entre rasgos sistémicos y periodontales siguió siendo limitada. De los dieciocho estudios del dominio de imagen diagnóstica, las redes convolucionales alcanzaron la mayor exactitud descrita para caries y cáncer oral, si bien pocos estudios usaron conjuntos públicos o validación externa. La evidencia respalda la viabilidad más que la preparación clínica; la heterogeneidad metodológica, las muestras pequeñas y la escasa validación externa fueron las limitaciones más frecuentes.
Artificial intelligence has become a pervasive analytical tool across the biomedical sciences, and its convergence with two rapidly maturing fields, the oral microbiome and dental genomics, has generated a dispersed body of studies that no single synthesis has yet consolidated. The oral cavity hosts one of the most densely sampled and best-characterized microbial communities of the human body, with hundreds of cultivable and not-yet-cultivable taxa whose composition shifts in caries, periodontal disease and oral squamous cell carcinoma.(1,2)
In parallel, genome-wide and transcriptome-wide studies have begun to resolve the heritable and developmental architecture of dental traits, from caries susceptibility to the genes that govern odontogenesis.(3,4) Machine learning offers a common computational language for both kinds of data, converting high-dimensional sequences, spectra and images into predictive or explanatory models.(5,6)
The clinical stakes are high, because the three domains converge on a set of conditions that account for a large share of oral disease: dental caries, periodontal disease and oral squamous cell carcinoma are among the most prevalent conditions of the human body, and their early detection remains the principal lever for reducing morbidity and cost. The oral microbiome is causally implicated in all three, through acidogenic communities in caries, dysbiotic biofilms in periodontitis and conserved taxa in carcinogenesis, while dental genomics resolves the host side of the same conditions, identifying the loci and developmental programmes that set individual susceptibility.
A tool that could learn from both the microbial and the host signals, and from the images that clinicians already acquire, would therefore address a genuine and persistent need rather than a contrived benchmark.
The appeal of applying machine learning to these fields lies in the structure of the data rather than in any single discovery. Sequencing produces thousands of taxa and genes per sample, hyperspectral and radiographic images produce millions of pixels per case, and the resulting matrices are too large for conventional univariate statistics to interrogate fully.(9,10) Supervised learners have been applied to predict clinical status from microbial or genomic features, while unsupervised and explainable methods have been used to identify conserved taxa and candidate variants.(7,11)
The result is a literature that spans three distinct subfields, oral microbiome modelling, dental genomics and clinical imaging, each with its own reporting conventions and its own limitations. The three domains differ in the data they feed to the models, and those differences shape what a model can plausibly claim. Oral microbiome studies rely on 16S ribosomal RNA amplicon sequencing, shotgun metagenomics or metabolomic profiles, each of which resolves a different layer of the community, from taxonomy to function to small-molecule output.(9,10)
Dental genomics studies rely on genome-wide association summary statistics, whole-exome or whole-genome data, and transcriptomic profiles of developing dental tissues.
The same algorithmic family, therefore, operates on inputs as different as a taxon count table and a panoramic radiograph, and the validity of a performance claim must be judged relative to the input type, not in the abstract.
The distinction between prediction and explanation is central to reading this literature, and it has been articulated most clearly in microbial population genomics.
This distinction explains many apparent inconsistencies in the field: imaging studies tend to report high accuracy because the task is narrowly perceptual, whereas genomic studies tend to report conservative, threshold-corrected associations because the task is inferential. Any synthesis that fails to separate these two modes risks comparing incommensurable results, and the present matrix is organized explicitly to avoid that error. The existing reviews in this area have generally addressed a single data type, most often radiographs or microbial biomarkers, without comparing the evidence across the microbiome, genomics and imaging domains.(16,17)
No prior synthesis, to the best of the present search, has organized the three domains side by side and weighted each by the quality of its evidence. The objective of this narrative review is threefold: first, to describe the applications of artificial intelligence to oral microbiome analysis; second, to describe the applications of artificial intelligence to dental genomics research; and third, to assess the quality and consistency of the evidence. The review addresses this objective through an evidence matrix that places each retained study in its domain and weights the synthesis by methodological quality.
Sources were searched in Scopus, PubMed and Crossref for publications on artificial intelligence, machine learning or deep learning in combination with oral microbiome, salivary microbiome, oral microbiota, dental genomics, genome-wide association or oral genome. The nucleus query in each source combined the artificial-intelligence terms with the oral terms using Boolean operators, and one hundred records were retrieved per source. The core queries paired the artificial-intelligence terms with the oral terms in the title, abstract and keyword fields, while supporting queries retrieved reviews on machine-learning approaches to microbiome classification and to genomics in oral disease for the introduction and discussion. The three record sets were merged and de-duplicated by digital object identifier, and the surviving records were screened against the inclusion criteria. When a source returned fewer than one hundred records, the full returned set was retained. Reference metadata were verified against Crossref, and Scopus coverage metadata was recorded where available to characterize the source base.
Studies were retained when they applied a machine-learning, deep-learning or generative method to oral microbiome, dental genomics or closely related oral diagnostic data, or when they constituted foundational reviews of the oral microbiome relevant to subsequent modelling. Studies that addressed non-oral systems, plant or animal genomics, or gut microbiome research without an oral component were excluded, as were purely technical records without a biological question. No date restriction was applied beyond the bibliographic window of 2018 to 2026, and the review remained narrative rather than systematic, so formal risk-of-bias scoring was not performed.
For each retained study, the design, the population or data source, the modelling approach, the principal finding and a quality note were extracted into an evidence matrix. Extraction was guided by the published abstract and, when available, by the structured methods and results of the source record. Findings were recorded at the level of discrimination, classification or association reported by each study rather than at the level of individual estimates, so that the matrix captured what each study demonstrated without reproducing every statistic. Studies were assigned to one of three domains, oral microbiome, dental genomics or clinical integration, on the basis of their primary data type. When the same study was retrieved from more than one source, the record with the fuller abstract was used for extraction, and the source of each study was retained for the source-coverage note. A study that combined imaging with genomics was assigned to the integration domain, so that the domain labels reflect the primary data type rather than a rigid taxonomy.
Quality was assessed descriptively across three dimensions: design, sample size and reproducibility. A study was rated higher when it used an external test set or public data, when it reported cross-validation or held-out performance, and when its population was described with sufficient detail to judge generalizability. The appraisal was used to weight the synthesis rather than to exclude studies, and it is reported alongside each domain in the development section. Studies were not excluded on the basis of quality, so that the full range of the literature remained visible, and the synthesis notes explicitly where a claim rests on a single low-quality study rather than on a pooled estimate. The appraisal also distinguished internal from external validity, because the two are frequently conflated in machine-learning reporting and a model that is accurate on its own training distribution is not necessarily accurate on a new one. Because the review is narrative, the appraisal was not reduced to a single score, and the quality notes in the matrix carry the judgment that a score would otherwise compress.
The synthesis proceeded in two steps. In the first step, studies were grouped by domain and counted, so that each claim about the literature could be anchored to a number of supporting studies. In the second step, the quality notes were used to weight the claims, giving precedence to systematic reviews, meta-analyses and multicentre designs over single-centre, retrospective reports. Where a claim rested on a single study, that fact was stated explicitly rather than generalized. The three objectives of the review map onto the three development subsections, and the conclusions restate them in the same order.
The evidence matrix summarizes the retained studies and anchors the synthesis that follows (Table 1). The synthesis proceeds domain by domain, and within each domain the statements are weighted by the number of studies that support them and by the quality notes in the matrix.
Table 1. Evidence matrix of studies applying artificial intelligence to the oral microbiome and dental genomics.
Study | Design | Population or data | Finding | Quality |
|---|---|---|---|---|
Cha et al. | Cross-sectional cohort | 182 Korean adults | XGBoost predicted chronological age from the salivary microbiome (MAE 7.49 years) | Exploratory, moderate sample |
Nakano et al. | ML modelling | Saliva samples | DL predicted oral malodour from salivary microbiota | Early proof of concept |
Zhang et al. | Longitudinal | OSCC progression | Dynamic changes in plaque and saliva microbiota across OSCC progression | Observational |
Tansirichaiya et al. | Comparative metagenomics | Five healthy adults | MetaPolyzyme lysis favoured Gram-positive detection and altered the resistome | Small methodological sample |
He et al. | Case-control with ML | 45 etomidate users, 44 controls | Oral dysbiosis with Actinomyces, Rothia and Atopobium; GLMBoost classified users | Case-control |
Kouznetsova et al. | Metabolomics with ML | Patient saliva | DL network distinguished oral cancer from periodontitis (79.54% accuracy) | Moderate accuracy |
Liu et al. | Hyperspectral imaging plus 16S | 153 subjects | Tongue-image model detected type 2 diabetes (ACC 0.897, AUC 0.803) | Cross-sectional |
Chen et al. | Generative multi-modal framework | In silico and in vivo | Neisseria flavescens identified as a geroprotective oral commensal | Preclinical |
Vishwakarma et al. | Systematic review and meta-analysis | Multi-cohort OSCC | Conserved taxa (Porphyromonas, Fusobacterium) enriched in OSCC | Multi-cohort synthesis |
Anghel et al. | Narrative review | Oral-gut-vascular axis | Periodontal dysbiosis linked to premature atherosclerosis | Mechanistic review |
Omidi et al. | Shared gene prioritization | DIAMANTE and FinnGen | 683 diabetes genes crossed threshold, none for periodontitis | Rigorous statistics |
Nogawa et al. | GWAS meta-analysis | Caries cohorts | Two novel loci associated with dental caries | Meta-analytic |
Chen et al. | RNA-seq of development | Murine dental epithelium and mesenchyme | 1281 and 1582 differential genes at two stages | Preclinical |
Yu et al. | Review | Oral potentially malignant disorders | Heterogeneous mechanisms and genomic instability across lesions | Conceptual review |
Saleh et al. | Transfer learning | 195 implants, two centres | Pre-trained CNNs predicted peri-implantitis from radiographs | Retrospective multicentre |
Cañedo Figueroa et al. | Systematic review | 45 studies | DL (CNN) in ~69%; sensitivity 0.85, specificity 0.90 for caries | Synthesis with meta-analysis |
Sarkar et al. | Generative AI with six ML models | 416 patients | GenAI automated preprocessing and periodontitis modelling | Retrospective |
Gill et al. | Comparative ML | Oral images | Early detection of oral disease with ML | Comparative |
Liang et al. | Bibliometric mapping | Literature | Trends of AI in oral disease diagnosis | Bibliometric |
Ghori et al. | Systematic review | Microbiome, biomarkers, radiology | AI in diagnosis and prediction across three data types | Broad synthesis |
Note: ML = machine learning; DL = deep learning; GWAS = genome-wide association study; OSCC = oral squamous cell carcinoma.
Of the fourteen studies that applied machine learning to the oral microbiome, eleven employed supervised classification or regression and three contributed foundational or single-cell accounts of the community itself. The supervised studies met their stated objectives with discrimination that ranged from modest to good, and the pattern of results was more consistent for continuous or binary outcomes than for multi-class discrimination. Age prediction was assessed in one cohort of 182 Korean adults, in which an XGBoost regressor achieved a mean absolute error of 7.49 years, an R-squared of 0.31 for chronological age, and an area under the curve of 0.76 when classifying participants by age band; an exploratory analysis of spousal cohabitation reported only a small unadjusted difference in community composition, which did not persist in a multivariable model that retained age and sex.(18)
Malodour prediction was reported in an early proof-of-concept study in which a deep network inferred oral malodour directly from salivary microbial composition, establishing the principle that odor-relevant taxa carry signal that a learned model can recover.
Disease-related shifts were reported across three studies, and of these the multi-cohort evidence was the strongest. A systematic review and meta-analysis that reprocessed raw 16S ribosomal sequences in QIIME 2 and applied LEfSe together with machine learning identified conserved taxa enriched in oral squamous cell carcinoma, including Porphyromonas, Fusobacterium, Campylobacter and Prevotella, with species such as Campylobacter showae and Prevotella loescheii among the discriminators.
A longitudinal account tracked dynamic changes in dental plaque and saliva microbiota across the progression of oral squamous cell carcinoma, and a mechanistic review linked periodontal dysbiosis to premature atherosclerosis through microbial translocation, endotoxemia, TLR2/TLR4 signalling and TMAO metabolism, naming hsCRP, IL-6, LPS, LBP, TMAO and oxidized LDL as candidate translational markers.(20,21) Of the three, only the first provided pooled quantitative estimates, and the other two contributed trajectory and mechanism rather than discrimination.
Substance-use and metabolite-based classification contributed two further applications, and both illustrated the value of the microbiome as a diagnostic substrate. Among users of etomidate, one case-control study of forty-five users and forty-four controls observed reduced species richness and diversity and a proliferation of Actinomyces, Rothia and Atopobium, with functional prediction by PICRUSt2 indicating enhanced carbohydrate and amino-acid metabolism and reduced network complexity; twelve machine-learning models were compared, and the GLMBoost ensemble discriminated users from controls.(22)
A separate metabolomic study distinguished oral cancer from periodontitis with a deep network built in TensorFlow that reached 79.54 percent accuracy under ten-fold cross-validation, a level that logistic regression and stochastic gradient descent matched but did not exceed.(23)
When restricted to systemic disease, one hyperspectral study integrated tongue imaging with 16S sequencing of fifty-three tongue-coating samples and built a type 2 diabetes classifier with an accuracy of 0.897 and an area under the curve of 0.803, correlating six genera with the imaging signal.(24) Of the fourteen studies, these supervised applications were rated as feasible but not clinically mature, because most relied on single-centre samples and none reported prospective external validation.
The modelling approaches themselves clustered into three families, and the choice of family tracked the structure of the data rather than any consensus on best practice. Gradient-boosted trees were used where tabular microbial abundances were the input, as in the age and substance-use studies, and they offered the practical advantage of ranking feature importance.(18,22) Deep networks were used where the input was sequence or image data with spatial or sequential structure, as in the malodour, metabolomic and hyperspectral studies.
Ensemble and multi-model comparisons were reported where the aim was to establish a robust classifier, and in each case a single model was selected on internal performance rather than on external evidence.
The feature sets that fed these models also differed in a way that bears on interpretability. Relative-abundance tables, which were the most common input, identify which taxa carry signal but say little about mechanism; functional predictions such as those from PICRUSt2 and metabolomic profiles move closer to mechanism but add their own estimation error; and imaging-derived features are the most information-dense but the least taxonomically interpretable.
Of the fourteen studies, only a minority reported which features drove the prediction, and of those that did, the drivers were biologically plausible, such as the proliferation of disease-associated taxa in the substance-use study.
Two studies contributed methodological or multi-omics advances rather than clinical prediction, and a further three supplied the biological context. A comparative metagenomic study assessed the effect of MetaPolyzyme lysis on five saliva samples and reported that chemical lysis favoured Gram-positive taxa, especially Streptococcus, over Gram-negative counterparts, while also altering the detectable resistome.
A generative multi-modality framework named AURORA built aging clocks, screened for age-gap-reducing interventions in silico and identified Neisseria flavescens as a geroprotective oral commensal, a prediction validated by lifespan extension in Caenorhabditis elegans and by restoration of the serum metabolome and gut microbiome toward younger states in aged mice.(25) The foundational accounts described the cultivation state of the oral microbiome, its manipulability and the effect of diet on its composition, and they framed the modelling studies by supplying the biological rationale for feature selection.
Across the fourteen studies, several methodological patterns were consistent enough to be treated as findings in their own right. Of the eleven supervised studies, eight reported a single performance estimate from an internal split or cross-validation, and only the multi-cohort systematic review pooled estimates across independent datasets.
When restricted to studies that disclosed their validation procedure, all used a train-test split or cross-validation, and none reported a fully independent, pre-registered external cohort. This absence of prospective validation, more than any single estimate, is what separates feasibility from clinical readiness in the oral-microbiome domain.
Of the seven studies in the dental-genomics domain, three employed genome-wide or transcriptome-wide association designs, two addressed the genomics of oral potentially malignant disorders, and two integrated omics into precision periodontology. The association studies met their objectives with varying degrees of shared signal, and the variation itself was informative. A genome-wide association meta-analysis identified two novel loci associated with dental caries, extending the list of reproducible caries-associated variants beyond the previously reported loci.
A transcriptome-wide study of murine dental epithelium and mesenchyme reported 1,281 and 1,582 differentially expressed genes at embryonic days 11.5 and 13.5, respectively, with enrichment of extracellular-region and extracellular-matrix-receptor terms, and prioritized extracellular proteoglycans and their glycosaminoglycan chains as candidate odontogenic regulators.
When restricted to shared genetic architecture between systemic and oral traits, one prioritization study analysed 19,280 diabetes genes and 19,378 periodontitis genes with a whole-gene test accounting for linkage disequilibrium, and it crossed 683 diabetes genes above a corrected threshold while none crossed for periodontitis; among published periodontitis entries, FCER1G and two SIGLEC5 variants met a relaxed comparison cutoff, and annotation-based elastic-net models retained dependence on locus-to-gene scores.
The genomics of malignant transformation contributed two reviews rather than new association data. One account described heterogeneous mechanisms, genomic instability and field cancerization across oral leukoplakia, proliferative verrucous leukoplakia and oral erythroplakia, and noted that these lesions share chronic inflammation and immune dysregulation yet follow divergent malignant trajectories.(29)
A companion account covered diagnostic adjuncts and biopsy techniques for these disorders, noting that no adjunct replaces histopathological examination.(30) The two precision-periodontology accounts synthesized molecular biomarkers, omics technologies, microbiome profiling and digital imaging as routes toward periodontal diagnosis and therapy.(31,32) Of the seven studies, the genomic findings were the most rigorously statistically grounded, but their translation toward clinical prediction remained indirect, because association does not itself produce a validated classifier.
A notable feature of the dental-genomics domain is the gradual shift from univariate association testing toward machine-learning-based prioritization, and the studies span that transition. The caries meta-analysis and the developmental transcriptome belong to the association tradition, in which loci and genes are identified by threshold crossing, whereas the shared-gene prioritization and the annotation-based elastic-net models belong to the predictive tradition, in which the goal is to rank candidates by their contribution to a model.
Of the eighteen studies in the diagnostic-imaging and clinical-integration domain, convolutional networks dominated, and caries and oral cancer were the two outcomes modelled in the largest number of studies, with three and four studies, respectively. Caries detection was assessed in three studies, of which a systematic review of forty-five studies reported that deep learning, chiefly convolutional networks, accounted for roughly sixty-nine percent of approaches, with the YOLO family, U-Net variants and transfer classifiers such as ResNet, VGG, DenseNet, EfficientNet and MobileNet as the most common architectures; reported accuracy ranged from 73.3 to 98.6 percent, pooled sensitivity was 0.85 and specificity 0.90, and only twelve percent of studies used public datasets.
Oral cancer detection spanned four studies, including a systematic review of imaging and an optimized convolutional model, both of which reported high accuracy on internal data, and a head-and-neck oncology account that positioned the models within the wider screening pathway.
Peri-implantitis prediction contributed one multicentre study in which five pre-trained convolutional architectures, ResNet50, DenseNet, VGG16, InceptionV3 and EfficientNet, were combined with four transfer-learning strategies and applied to 195 implants across two centres with patient-wise splits to avoid leakage.(39) Periodontitis risk stratification was addressed by one generative-artificial-intelligence workflow that automated preprocessing, correlation analysis and the development of six models, logistic regression, decision tree, random forest, gradient boosting, support vector machine and k-nearest neighbours, in 416 patients, with five-fold cross-validation and McNemar's test.
Among high-quality studies, the clearest signal was that deep learning achieved strong internal discrimination but weak external generalizability, a pattern driven by the scarcity of shared datasets.
The imaging tasks themselves were heterogeneous, and the differences matter for how the reported accuracy should be read. Detection tasks asked whether a lesion or a carious surface was present, segmentation tasks asked where it was located, and classification tasks asked what it was, and the reviewed studies mixed the three without a consistent reporting convention.(13,37)
Transfer learning from ImageNet-scale pre-training was the dominant strategy for small dental datasets, and its benefit was clearest in the peri-implantitis study, where patient-wise splits guarded against leakage across the two centres.(39) Of the eighteen studies, those that reported sensitivity and specificity rather than accuracy alone were the exception, and the absence of a shared metric made direct comparison across the domain difficult.
Three broad syntheses mapped the field and described artificial intelligence across microbiome, biomarker and radiological data in oral disease, and their agreement on the dominance of supervised imaging is notable.
A final pattern in the integration domain concerns reproducibility, and it is the pattern with the largest practical consequence. Of the eighteen studies, a minority released public datasets or pre-trained models, and the systematic review that quantified this reported that only twelve percent of caries-detection studies used public data.
The consequence is that most reported accuracies are not independently verifiable, and the field's claim to generalizability rests on internal rather than external evidence. When restricted to the studies that did use public data, the discrimination estimates were generally lower and more heterogeneous, which suggests that the higher figures in the remainder reflect, at least in part, the favourable conditions of private datasets. The generative workflow for periodontitis illustrates the same point from the opposite direction: its value lay in reducing the engineering burden, not in raising discrimination, and the authors reported only modest predictive power.
The principal finding of this matrix is that artificial intelligence in oral microbiome and dental genomics research has demonstrated feasibility rather than clinical readiness, and that the three domains differ sharply in how much their reported performance can be trusted. Supervised models of the oral microbiome produced modest to good discrimination for age, malodour, substance use and disease status, but the estimates came almost entirely from single-centre samples without prospective validation.
The dental-genomics studies were the most statistically rigorous, yet none of them yielded a classifier, and the shared prioritization between systemic and periodontal traits was largely absent. The imaging studies reported the highest accuracy, but that accuracy was concentrated in internal test sets, and only a minority of them released public data. This three-way split, more than any individual result, is the contribution of the present synthesis.
The oral-microbiome findings agree with the broader microbiome literature in which supervised learning has repeatedly discriminated disease from health but has struggled to generalize across cohorts. The conserved taxa enriched in oral squamous cell carcinoma, including Porphyromonas and Fusobacterium, parallel the enrichment patterns reported in gut and lung studies in which machine learning identified candidate taxa but the models degraded on new data.(7,47) The age-prediction result in the Korean cohort is consistent with the observation that salivary microbial composition carries chronological signal, although the modest variance explained indicates that microbial age is not yet a precise biomarker.
The drug-depletion and autism-associated gut models, although non-oral, corroborate the general point that microbial features support classification when the phenotype is biologically proximal to the community.(48,49) Studies that assessed explainable models in genome-wide association settings reported the same tension seen here: predictive performance and biological interpretability rarely peak together.
The dental-genomics findings diverge from the imaging findings in an instructive way. Where imaging models reported near-ceiling accuracy, the genomic studies reported conservative, threshold-corrected results, and the absence of shared signal between diabetes and periodontitis genes is a caution against assuming that systemic and oral traits share genetic architecture.
The two novel caries loci and the developmental odontogenic genes are reproducible contributions, but they are discoveries about biology rather than tools for prediction.(3,4) The distinction between explanatory and predictive modelling, articulated in the microbial population-genomics literature, applies with particular force here: a locus that survives a genome-wide threshold may explain little variance at the level of an individual patient.(5,15) The methodological reviews of genomic deep learning reinforce this point by showing that noncoding-variant and sequence-to-activity models optimize molecular prediction, not clinical outcome.
The imaging studies, in turn, are best interpreted against the literature on single-cell and imaging genetics, in which self-supervised and deep representations have improved prediction but have also concentrated the field on narrow benchmark tasks.
The multi-omics and screening reviews in the gut domain show the same pattern, in which methodological sophistication outpaces validation, and they suggest that the oral field is following a trajectory already visible elsewhere.(9,53) The oral health policy literature makes a related point: the value of these tools depends less on algorithmic novelty than on whether they can be embedded in prevention-oriented, equity-conscious care pathways.
The consistency of the findings across the three domains can be summarized in a single sentence: the stronger the statistical control, the weaker the clinical claim, and the stronger the clinical claim, the weaker the statistical control. This pattern held within each domain as well as between them. Among the microbiome studies, the multi-cohort systematic review offered the most robust pooled estimates but the least clinically actionable output, whereas the single-centre classifiers offered the most actionable output with the least robust evidence.(7,22)
Among the genomic studies, the threshold-corrected associations were the most reproducible but the least predictive, and among the imaging studies the highest accuracies came from the least externally validated designs.(12,13) A single study that combined threshold-corrected discovery, validated discrimination and functional context would be the natural next step, and the multi-omics templates already developed for the gut describe exactly that combination.
Placed side by side, the three domains reveal an inverse relationship between the strength of the statistics and the clinical immediacy of the output. The genomic studies had the strongest inference and the least immediate clinical product, the imaging studies had the weakest generalizability and the most immediate product, and the microbiome studies occupied the middle ground. This ordering is not a defect of any single study but a property of the underlying tasks: genome-wide inference is designed to control false positives, image classification is designed to maximize accuracy on a perceptual task, and microbiome modelling is constrained by the compositional and batch-sensitive nature of its input.(5,15)
A mature translational pipeline would need the strengths of all three, threshold-corrected discovery from genomics, validated discrimination from imaging and functional context from the microbiome, and none of the reviewed studies yet combines them in a single design. The multi-omics reviews that describe such integration in the gut demonstrate that the combination is technically feasible, and they provide a template that the oral field has not yet adopted at scale.
A second cross-cutting observation is that several families of methods that are now standard in the wider genomics and imaging literature were largely absent from the oral studies reviewed here. Self-supervised learning, which builds representations directly from unlabelled images and has reshaped imaging genetics, appeared in none of the oral microbiome or imaging studies, whose models remained predominantly supervised.
Explainable artificial intelligence, which converts a deep model into an account of the variants or taxa that drive it, was likewise rare outside the genome-wide setting, and its absence limits the clinical trust that a black-box classifier can command.
Single-cell approaches, which have resolved cell-state plasticity in cancer, offer a complementary route to the developmental transcriptomes of odontogenesis and have not yet been combined with machine learning in the dental domain.
The implications for practice and for research follow directly from these patterns. For practice, the strongest near-term use is decision support in imaging, where caries and oral cancer models already reach clinically meaningful sensitivity and specificity on internal data, provided that local validation is performed before deployment.(13,36) The human-interface literature cautions that such support is most valuable when the model is kept in a supporting role, with the clinician as the final arbiter rather than a passive recipient of a score.
For research, the clearest priorities are public datasets, pre-registered external validation and reporting standards that separate internal from external performance, and these are concrete steps rather than an open-ended call for more data. The oral health policy literature reinforces the point that such tools will deliver value only if they are embedded in prevention-oriented, equity-conscious pathways rather than layered onto intervention-based care.
The limitations of this review fall into three groups: those of the included studies, those of the evidence base as a whole, and those of the review process itself. The included studies were limited by small samples, retrospective or cross-sectional designs, and a near-universal absence of prospective external validation; each of these limits the generalizability of the reported performance, and a recommendation follows each. Small samples inflate the variance of performance estimates, so future work should report confidence intervals and power calculations rather than point accuracy alone. The scarcity of external validation means that high internal accuracy cannot be distinguished from overfitting, so shared benchmark datasets and pre-registered test sets should become the default. The evidence base as a whole was limited by publication bias toward positive results and by the uneven availability of abstracts across sources, which restricted extraction for a minority of records; a systematic review with full-text retrieval would address this by recovering methods and results that titles and abstracts omit. The review process was limited by its narrative, non-systematic design and by the descriptive rather than quantitative appraisal of quality; a formal risk-of-bias tool and a meta-analysis of discrimination metrics would strengthen future syntheses.
The applications of artificial intelligence to oral microbiome analysis are dominated by supervised classification and regression that achieve modest to good discrimination for age, malodour, substance use and disease status but remain unvalidated beyond single-centre samples.
The applications of artificial intelligence to dental genomics research identify reproducible candidate loci and developmental genes but produce discovery rather than clinical prediction, with little shared signal between systemic and periodontal traits. The quality and consistency of the evidence are uneven across the three domains, with statistically rigorous genomics studies, high-accuracy but weakly generalizable imaging studies, and feasible but unvalidated microbiome models.
No financing
None
Drafting – original draft: Ana María Chaves Cano.
Writing–review and editing: Ana María Chaves Cano.