Modular Deep-Learning Pipelines for Dental Caries Data Streams: A Twin-Cohort Proof-of-Concept.

Abstract

Background: Dental caries arise from a multifactorial interplay between microbial dysbiosis, host immune responses, and enamel degradation visible on radiographs. Deep learning excels in image-based caries detection; however, integrative analyses that combine radiographic, microbiome, and transcriptomic data remain rare because public cohorts are seldom aligned. Objective: To determine whether three independent deep-learning pipelines-radiographic segmentation, microbiome regression, and transcriptome regression-can be reproducible implemented on non-aligned datasets, and to demonstrate the feasibility of estimating microbiome heritability in a matched twin cohort. Methods: (i) A U-Net with ResNet-18 encoder was trained on 100 annotated panoramic radiographs to generate a continuous caries-severity score from a predicted lesion area. (ii) Feed-forward neural networks (FNNs) were trained on supragingival 16S rRNA profiles (81 samples, 750 taxa) and gingival transcriptomes (247 samples, 54,675 probes) using randomly permuted severity scores as synthetic targets to stress-test preprocessing, training, and SHAP-based interpretability. (iii) In 49 monozygotic and 50 dizygotic twin pairs (n = 198), Bray-Curtis dissimilarity quantified microbial heritability, and an FNN was trained to predict recorded TotalCaries counts. Results: The U-Net achieved IoU = 0.564 (95% CI 0.535-0.594), precision = 0.624 (95% CI 0.583-0.667), recall = 0.877 (95% CI 0.827-0.918), and correlated with manual severity scores (r = 0.62, p < 0.01). The synthetic-target FNNs converged consistently but-as intended-showed no predictive power (R² ≈ -0.15 microbiome; -0.18 transcriptome). Twin analysis revealed greater microbiome similarity in monozygotic versus dizygotic pairs (0.475 ± 0.107 vs. 0.557 ± 0.117; p = 0.0005) and a modest correlation between salivary features and caries burden (r = 0.25). Conclusions: Modular deep-learning pipelines remain computationally robust and interpretable on non-aligned datasets; radiographic severity provides a transferable quantitative anchor. Twin-cohort findings confirm heritable patterns in the oral microbiome and outline a pathway toward future clinical translation once patient-matched multi-omics are available. This framework establishes a scalable, reproducible foundation for integrative caries research.