Abstract
The precise control of cell fate is driven by a hierarchical regulatory network (HRNet) where transcription factors (TFs) and cis-regulatory elements (CREs) orchestrate the expression of target genes (TGs) through complex causal actions. While single-cell multi-omics technologies provide multi-dimensional data to resolve regulatory networks, existing methods often fail to capture their hierarchical and causal properties. We propose SMOGT (Single-cell Multi-Omics Graph Transformer), a graph representation learning method to decipher HRNet. SMOGT embeds epigenetic mechanism into Heterogeneous Graph Transformer (HGT) by structuring information flow along a hierarchical-guided meta-path (TF-TF → TF-CRE → CRE-CRE → CRE-TG), and employs a semi-supervised strategy to ensure network accuracy. Validated against ChIP-seq and HiC-seq benchmarked datasets, SMOGT showed significantly higher accuracy in predicting transcriptional regulation (TF-CRE) and long-range chromatin conformation (CRE-CRE). The HRNet scaffolds downstream modules that mechanistically link network architecture to cell fate. The multi-layer random walk (MRWR) module identifies driver regulators and their TGs. The BioStreamNet module predicts shifts in cell fate trajectories following in silico perturbations within gene-specific HRNet formed by extracting regulatory weights during TG expression prediction. In hematopoietic stem cell differentiation, SMOGT elucidated the hierarchical causal cascade from driver TFs that governs lineage commitment. In melanoma epithelial-to-mesenchymal transition (EMT), it revealed a critical therapeutic window for reversing the process, and in Acute Myeloid Leukemia (AML), it uncovered hub-CREs with significant prognostic value. By accurately modeling hierarchical causality, SMOGT provides a robust tool to dissect and predict cell fate dynamics in both development and disease.