Four datasets used in this work.

Understanding gene regulation during cell differentiation requires effective integration of multi-omics single-cell data. In this study, we propose BranchKGN, a heterogeneous graph transformer-based framework for identifying branch-specific key genes along cell differentiation trajectories. By integrating scRNA-seq and scATAC-seq data into a unified gene representation, we infer differentiation trajectories using Slingshot and construct a heterogeneous graph capturing gene–cell relationships. Through attention-based graph learning, BranchKGN assigns gene importance scores within each cell, enabling the identification of genes consistently informative across branch point cells and their descendant lineages. These genes are then used to reconstruct gene regulatory networks and differentiation trajectories. Validation on three independent datasets demonstrates that the identified gene sets not only capture key regulators of cell fate bifurcation but also support accurate reconstruction of differentiation trajectories. Our results highlight the effectiveness of BranchKGN in dissecting gene regulation dynamics during cellular transitions and provide a valuable tool for multi-omics single-cell analysis.