PerioDynaCausal-GT: a dynamic causalgraph transformer for single-cell-informed transcriptomic discovery ofimmuno-epigenetic drivers of periodontalattachment loss.

Abstract

Introduction

Progressive periodontal attachment loss (PAL) signifies the final and irreversible stage of periodontitis. It results from the convergence of a dysbiotic microbial challenge, maladaptive innate immune escalation, and epigenetically reinforced stromal activation. Despite this understanding, the precise molecular hierarchy responsible for this destruction at a cellular resolution has remained undefined. We introduce PerioDynaCausal-GT, a Dynamic Causal Graph Transformer that synthesizes differential expression data, single-cell RNA sequencing deconvolution, and pathway-encoded causal graph learning. This system systematically ranks immuno-epigenetic drivers of PAL based on large-scale, multi-cohort transcriptomic evidence. 

Methods

Applying this framework to 19,177 gene-level features across two independent gingival biopsy cohorts—GSE10334 (n=247; 183 diseased, 64 healthy) and GSE16134 (n=310; 241 diseased, 69 healthy)—we identified 1,003 concordant differentially expressed genes (DEGs; 551 upregulated, 452 downregulated) with cross-cohort log₂FC concordance of Pearson r=0.973. Bindea single-sample GSEA across 23 immune populations yielded a cross-cohort immune infiltration concordance (Pearson r=0.995; 19/23 populations FDR<0.05 in both cohorts independently), establishing a reproducibility standard that far exceeds prior periodontal immunoprofiling studies.

Results

A four-component Driver Priority Score assessed a 94,108-cell gingival single-cell atlas, identifying CD79A, IL1B, and IRF4 as the foremost PAL effectors based on differential expression magnitude, validation, pathway load, and cell-type specificity. IL6 is ranked within the top ten (priority=9.40), primarily influenced by pathway burden (C_z=10.22; 30 pathways), despite a modest fold change, signifying its role as a significant pleiotropic regulator that may be undervalued if considering differential expression alone. Classifiers built on the 30-gene signature demonstrated high AUC values and low Brier scores when validated externally; furthermore, LM22 CIBERSORT analysis corroborated the enrichment of disease-associated T follicular helper cells (FDR=1.4×10⁻¹³), M2 macrophages (FDR=2.1×10⁻¹¹), and plasma cells (FDR=7.9×10⁻¹⁰).

Conclusion

Clinically, PerioDynaCausal-GT recognizes periodontal attachment loss as a reproducible immuno-stromal disorder governed by B-plasma cell proliferation, IL6/CXCL1 cytokine networks, and compromise of the epithelial barrier.