Federated Multi-Omics Learning Reveals a Conserved Epithelial Immune Commitment Axis Associated with Epithelial Immune Commitment in Periodontitis Across Distributed Cohorts

Abstract

Background: Periodontitis is a chronic immune-inflammatory disease driven by dysbiotic microbial challenge at the gingival epithelial interface. Despite substantial molecular profiling of diseased gingival tissue, no federated, privacy-preserving computational framework has identified the cross-cohort-conserved epithelial molecular axis that governs disease irreversibility. This study aimed to identify and validate a conserved epithelial immune commitment axis in periodontitis using biologically constrained federated graph learning across distributed transcriptomic cohorts.

Materials and Methods: Two large bulk RNA microarray datasets — GSE10334 (n = 247; GPL570) and GSE16134 (n = 310; GPL570) were treated as independent federated nodes. Within each node, quantile normalization, Welch-corrected differential expression with Benjamini–Hochberg false discovery rate control, and construction of a 190-gene epithelial-immune panel-based gene-gene graph (pathway co-membership + Spearman |ρ| ≥ 0.50 edges) preceded local graph convolutional network (GCN) training. A federated server aggregated local model weights using biologically constrained FedAvg (FedBio), weighting gene contributions by cross-node differential expression concordance. An Epithelial Commitment Score (ECS) was derived from the consensus differential expression signature. Single-cell validation was performed in GSE164241 (n = 18,142 post-QC cells; 1,029 epithelial cells; 6 gingival samples: 3 healthy, 3 periodontitis), with Leiden clustering (resolution 0.40), epithelial subcluster module scoring, and derivation of an Epithelial Inflammatory Commitment Axis (EICA = inflammatory score − barrier score).

Results: Cross-node differential expression concordance was exceptional (Spearman ρ = 0.946; panel ρ = 0.991; p < 1 × 10⁻³⁰⁰). The ECS achieved AUROC = 0.909 (Mann–Whitney U p = 1.2 × 10⁻²²) in GSE10334 and AUROC = 0.926 (p = 1.8 × 10⁻²⁷) in GSE16134. FedBio improved over plain FedAvg (AUROC: +0.031 and +0.035, respectively). Conventional classifiers trained per-node achieved higher within-node AUROCs (SVM: 0.906, 0.932; LR: 0.889, 0.932) and excellent cross-node generalization (AUROC 0.978–0.992). The local GCN underperformed on cross-node transfer (AUROC 0.672–0.686), indicating that graph-topological generalization requires larger federated networks. Single-cell analysis identified epithelial subclusters 6 (74% periodontitis-derived) and 8 (97% periodontitis-derived) with elevated NF-κB, pyroptosis, and IL-1/IL-6 module scores. Bulk ECS projection onto single epithelial cells correlated with the EICA (Spearman ρ = 0.516; p = 3.08 × 10⁻⁷¹).

Conclusions: This work identifies a conserved Epithelial Immune Commitment Axis in periodontitis, marked by barrier collapse (CDH1/EPCAM/OCLN) and inflammatory escalation (MMP7/IL-1β/ZBP1/CXCL). FedBio allows cross-institutional discovery without sharing raw data and produces a transparent ECS biomarker. Restoring barriers and suppressing inflammasomes are promising therapeutic strategies to halt periodontal breakdown.