Decoding Cell-type Specific Gene Regulation with CellReg3D by Integrating DNA Large Language Models and 3D Genome Structure

Gene regulation is a fundamental biological process that underpins organismal growth and development. Recent progress in artificial intelligence, complemented by expanding computational resources, has facilitated large-scale modeling of gene regulatory mechanisms. Despite the proliferation of models, accurately representing long-range interactions within the genome remains a formidable challenge.

In this study, we introduce CellReg3D, a sophisticated multi-scale and multi-modal deep learning framework aimed at modeling cell-type-specific gene regulation by merging local sequence patterns with long-range genomic interactions. Conventional methods typically disregard the impact of 3D genome architecture and struggle to synthesize sequence-based features with chromatin interaction data.

CellReg3D addresses these shortcomings by employing a DNA language model to encode sequence-level information and integrating Hi-C-derived 3D chromatin interactions to capture the spatial organization of chromatin. We show that CellReg3D proficiently learns epigenetic patterns and chromatin activity states, attaining state-of-the-art performance in predicting epigenetic marks and gene expression profiles across various cell types.

Through interpretability tools, we demonstrate the model's ability to discern cell-type-specific regulatory mechanisms and highlight functional non-coding variants. Moreover, we utilize CellReg3D to refine the interpretation of expression quantitative trait loci (eQTLs) and identify novel potential gene regulatory interactions. In summary, CellReg3D presents an integrative approach to deciphering regulatory genomics by harmonizing sequence data and 3D genome structure, thereby advancing variant prioritization, elucidating disease mechanisms, and facilitating personalized functional genomics.