Transformer-Based Deep Learning for the Joint Prediction of Enhancer-Promoter Interactions in Single-Cell Chromatin Accessibility Profiles

Deciphering the complex regulatory network of enhancer-promoter interactions (EPIs) is fundamental to understanding gene expression dynamics across heterogeneous cell populations. Traditional computational models for EPI prediction rely heavily on bulk epigenomic data, which obscures cell-type-specific regulatory landscapes. While single-cell chromatin accessibility profiling (scATAC-seq) offers unprecedented resolution, its inherent sparsity and high noise levels present substantial computational challenges. To address these limitations, we present EpigeneticTransformer (EpiTrans), a novel deep learning framework that leverages transformer-based self-attention mechanisms to jointly predict EPIs directly from sparse single-cell chromatin accessibility profiles. EpiTrans models long-range genomic dependencies by treating genomic bins as tokens, learning contextualized representations of enhancers and promoters simultaneously. We evaluated EpiTrans on benchmark scATAC-seq datasets from diverse human tissue types. Our model demonstrated superior performance over existing state-of-the-art methods, achieving an area under the precision-recall curve (AUPRC) of 0.84 and robustly generalizing across unseen cell types. Furthermore, attention weight analysis revealed that EpiTrans captures biophysically meaningful interactions corresponding to known loop-extruding CTCF binding sites and active chromatin marks. This work underscores the power of transformer architectures in resolving sparse single-cell genomic data, providing a scalable and interpretable tool for dissecting cell-type-specific gene regulation in development and disease.