A Graph Contrastive Learning Analysis of Gene Expression Networks in Orthodontic Tooth Movement and Periodontal Inflammation

A BSTRACT Introduction: Orthodontic tooth movement (OTM) and periodontal inflammation involve periodontal tissue remodeling through inflammatory and biological responses. The underlying gene expression networks and their similarities/differences, remain incompletely understood. This study developed a graph contrastive learning (GCL) model to identify shared and condition-specific gene network features in OTM and periodontal inflammation. Materials and Methods: Gene coexpression networks were built from 250 genes under orthodontic force and periodontal disease. A graph neural network-based GCL model with three convolutional layers and a projection head learned embeddings that captured network structures across conditions (final loss ~3.656). Hub genes were identified, overlaps assessed using Fisher’s exact test; effect sizes between hubs and nonhubs were measured with Cohen’s d . Machine learning classifiers predicted differential expression and hub status, evaluated by area under the curve (AUC). Uniform manifold approximation and projection (UMAP) visualized embeddings and network topology. Results: The orthodontic network (111 genes and 105 edges) was larger, more connected than the periodontal network (81 genes and 28 edges). OTM-specific hubs (24) included CTNNBIP1 and ENO1 , while periodontal-specific hubs ( CHD5 , DFFB , and GABRD ) were few and mostly overlapped with OTM. Periodontal-specific hubs ( CHD5 , DFFB , and GABRD ) were few, and although most periodontal hubs overlapped with OTM, these specific genes represented a small subset uniquely enriched in periodontal inflammation. The GCL model’s 32-dimensional embeddings captured both shared and distinct modules, visualized as UMAP clusters. Moderately expressed genes showed slightly higher significance (4%), but only 6% remained significant after correction. Classifiers performed strongly (AUC: 0.994 ± 0.013 for differential expression and AUC: 0.973 ± 0.055 for hub prediction). Conclusion: OTM and periodontal inflammation share many genes and pathways, but differ in network topology and hubs. GCL distinguished shared and specific features, providing insights into tissue remodeling and potential therapeutic targets in comparative network biology.