EpiNet: A multilayer integrative atlas of the human epigenetic regulatory network reveals principles of complex assembly, domain modularity, and lncRNA scaffolding

Background EpiFactors is the reference manually curated database for human epigenetic regulators (v2.1, September 2024), cataloguing 801 proteins, 73 macromolecular complexes and 124 long non-coding RNAs (lncRNAs). Despite its widespread use as a gene-set resource, EpiFactors has never been analysed as an integrated multilayer system. We hypothesised that a network representation would reveal organisational principles—modular complex assembly, recurrent domain architectures, and lncRNA scaffolding—that remain invisible to list-based approaches. Methods We constructed EpiNet, a three-layer integrative atlas. Layer A is a protein–complex bipartite graph (membership edges, n = 423, m = 570). Layer B is a protein–Pfam domain bipartite graph (n = 1,467, m = 1,757). Layer C is an lncRNA–target bipartite graph inferred by conservative regular-expression text mining of the EpiFactors lncRNA annotation table (n = 210, m = 207). For each layer we computed degree distributions (power-law fitting), density, clustering and connected components. Community structure was detected with the Leiden algorithm (RBConfigurationVertexPartition) across seven resolution parameters (γ = 0.20–1.50), calibrated by gap statistic and Normalized Mutual Information stability. Multilayer protein centrality was computed as a weighted sum of z-standardised degree, betweenness and eigenvector scores across Layers A and B. Hub status (known epigenetic regulators vs. others) was predicted via 10-fold stratified cross-validation using regularised Random Forest (max_depth = 5, min_samples_leaf = 5, min_samples_split = 12, max_features='sqrt') and logistic regression, with overfitting diagnosed by learning curves, and performance evaluated by ROC-AUC, PR-AUC, F1, MCC and Brier score with bootstrap confidence intervals (n = 1,000). Permutation testing (n = 200) and hold-out validation (80/20 split) confirmed that predictions exceeded the null distribution. Pfam domain co-occurrence was tested for enrichment with Fisher's exact test and Benjamini–Hochberg FDR correction. Robustness was assessed by comparing observed metrics against 100 degree-preserving randomised networks. Results Layer A exhibited scale-free-like behaviour (power-law α = 2.62, x_min = 3) with core density 6.48 × 10⁻³. Layer B was larger and sparser (α = 2.86, x_min = 8). The filtered Layer C contained 127 edges among 62 nodes (density = 0.067). Leiden detection at calibrated resolutions yielded 30 communities in Layer A (γ = 1.25, Q = 0.834), 207 communities in Layer B (γ = 1.00, Q = 0.916), and 12 communities in Layer C (γ = 1.50, Q = 0.437). Multiplex centrality ranked WDR5 as the top super-hub, followed by ACTL6A/BAF53A, HDAC1, ASH1L and SMARCC1/BAF155. Domain-combinatorial analysis identified PF00271 + PF00176 (SNF2-related N-terminal and helicase domains) as the most significantly enriched signature of chromatin remodellers (q = 4.01 × 10⁻⁴⁴). Layer C analysis revealed BAALC-AS1 as the master scaffold (degree = 14), together with MEG3, NEAT1 and XIST. Degree-preserving randomisation confirmed that observed clustering and modularity were non-random (empirical p < 0.001 for all layers).Hub classification achieved ROC-AUC = 0.89 (95% CI: 0.82–0.95) with permutation p < 0.001, confirming that multiplex centrality captures genuine biological signal rather than overfitting. Conclusion EpiNet demonstrates that the human epigenetic machinery is organised as a hierarchical, modular and non-random multilayer system. WDR5 emerges as a structural bridge linking histone methylation, acetylation and remodelling complexes, while recurrent Pfam dyads and lncRNA scaffolds provide a combinatorial grammar for complex assembly. The atlas and all derived tables are released as a community resource for hypothesis generation in epigenetic systems biology.