Abstract
Cancer-associated epigenetic remodeling involves a complex interplay between DNA secondary structures and covalent cytosine modifications, yet predicting endogenously folded, chromatin-constrained G-quadruplex (G4) dynamics remains challenging. Here, we develop G4SCOPE, a multi-task deep learning framework that models chromatin-associated G4 occupancy and reconstructs quantitative high-resolution CUT&Tag profiles directly from primary DNA sequence alone. We initially trained G4SCOPE on genotype-stratified G4 enriched cut&tag landscapes generated across an isogenic series of human breast epithelial cells (WT, TP53−/−, TP53−/−;BRCA1−/−, and TP53−/−;BRCA2−/−). G4SCOPE integrates convolutional motif learning with transformer self-attention to decode long-range sequence contexts. The model robustly distinguishes chromatin-associated G4s from GC-matched loci with intrinsic in vitro folding potential, generalizes across distinct breast lineages, and transfers accurately to an independent K562 cellular context. Strikingly, genotype-resolved mapping reveals a profound divergence in genome maintenance states: BRCA1 deficiency indicated a global inflation of chromatin-embedded G4 density, whereas BRCA2 loss was associated with systemic, chromosome-wide structural depletion. Furthermore, unsupervised clustering of model-derived latent sequence representations identified distinct biological G4 subclasses that occupy unique regulatory class genomic environments associated with distinct local 5-methylcytosine and 5-hydroxymethylcytosine landscapes. Together, our framework decodes the primary sequence grammar governing cell-based G4 occupancy and demonstrates that sequence-defined G4 topologies harbor intrinsic and distinct local epigenetic features.