NeuroG4Net: A Disease-Aware Hierarchical Multi-Attention Deep Learning Framework for G-Quadruplex Prediction in Neurodegenerative Disease Genomes Integrating H3K27ac Enhancer Marks and ATAC-seq Epigenomic Signals

G-quadruplex (G4) DNA structures forming at neurodegeneration risk-gene promoters are increasingly recognised as epigenomic regulators of transcriptional dysregulation in Alzheimer’s disease (AD), amyotrophic lateral sclerosis (ALS), and Parkinson’s disease (PD). Yet no existing computational G4 prediction framework has been designed specifically for neurodegenerative applications, and none incorporates the disease-relevant H3K27ac active-enhancer marks and chromatin accessibility signals that define the neuronal epigenomic landscape. We address this gap with NeuroG4Net, a disease-aware hierarchical multi-attention deep learning framework that integrates DNA sequence, H3K27ac ChIP-seq, and ATAC-seq signals from 17 patient-derived postmortem brain tissue samples spanning the three diseases. Training proceeds in two phases: pre-training on Zenodo 8144456 (30 epochs, 716,310 windows) followed by fine-tuning on 58,754 patient-derived BG4 ChIP-seq G4 peaks drawn from AD, ALS, and PD cohorts. The architecture combines a three-block ResNet 1D-CNN sequence encoder, a dual epigenomic encoder, Cross-Modal Attention (CMA), and a Hierarchical Multi-head Attention (HMA) mechanism operating at four biologically motivated genomic scales, trained jointly with a disease-contrastive multi-task loss. On five-fold cross-validation, NeuroG4Net reaches AUPRC = 0.934 ± 0.002, AUROC = 0.963 ± 0.003, Accuracy = 94.8 ± 0.21%, F1 = 93.7 ± 0.23%, and MCC = 0.882 ± 0.004, outperforming nine competitive baselines with Wilcoxon signed-rank significance (p < 0.001; effect sizes 0.498–0.921). An eight-configuration ablation confirms H3K27ac as the single most informative epigenomic predictor (|SHAP| = 0.203 ± 0.041, rank 1), and disease-specific G4 analysis identifies 2,341 AD-specific, 1,983 ALS-specific, and 1,542 PD-specific G4 loci enriched at GWAS risk-gene promoters (OR = 3.28–5.63; all p < 10⁻¹¹). Integrated Gradients confirms 90.7% G-run positional concordance. Taken together, these results establish NeuroG4Net as the first disease-aware computational G4 epigenomics framework for neurodegeneration and yield a prioritised therapeutic G4 target catalogue.