EpiRG4Net: An Epitranscriptomic-Aware Deep Learning Framework for In Vivo RNA G-Quadruplex Prediction Integrating m6A Methylation and Chromatin Accessibility Signals

RNA G-quadruplexes (rG4s) are non-canonical secondary structures that regulate mRNA translation, stability, and splicing in a cell-context-dependent manner. Their formation in living cells depends not only on sequence G-richness but also on N6-methyladenosine (m6A) modification and co-transcriptional chromatin accessibility — regulatory layers that existing computational rG4 prediction methods have overlooked entirely. We present EpiRG4Net, the first epitranscriptomic-aware deep learning framework to integrate RNA sequence (one-hot + k-mer), m6A methylation density profiles from MeRIP-seq, and ATAC-seq chromatin accessibility signals, using a three-block ResNet 1D-CNN architecture combined with Cross-Modal Attention (CMA, h=8), Spatial Pyramid Attention (SPA, scales S={4,8,16,32}), a BiLSTM layer, and a three-head multi-task output. Trained on 328,420 curated 150 nt rG4-seq windows from HeLa and MCF7 in vivo DMS-MaPseq experiments, EpiRG4Net achieves AUPRC=0.923±0.003, AUROC=0.960±0.005, Accuracy=94.2±0.27%, F1=93.1±0.38%, MCC=0.871±0.005, and Specificity=94.0% on five-fold cross-validation, outperforming nine baseline methods with Wilcoxon signed-rank significance (p<0.001; effect sizes r=0.531–0.891; ΔAUROC=+0.030 to +0.130). An eight-configuration ablation confirms that each component contributes independently (total gain +8.3 AUPRC points). Integrated Gradients achieves 91.3% G-run positional concordance (χ²p<10⁻¹⁶), and SHAP identifies m6A peak proximity as the second-ranked predictor (|SHAP|=0.142±0.031). Model complexity analysis confirms 2.4 ms inference latency, enabling whole-transcriptome screening. Differential rG4 analysis identifies 2,156 HeLa-specific and 1,874 MCF7-specific rG4 loci, with biologically coherent GO enrichment at oncogenic transcripts.