spaNetReg: A Dependency-Aware Graph Neural Network for Inferring TRNS from Spatial ATAC-Seq

Transcription factor regulatory networks (TRNs) are central to the regulation of gene expression, development, and pathogenesis. The emergence of spatial omics technologies has created new opportunities for systematically decoding TRNs; however, the characteristically high noise and sparsity hinder accurate inference by existing computational methods. To address this, the inherent spatial dependencies within biological systems can be leveraged as an informative prior to enhance the robustness of TRN inference. Here, we developed spaNetReg, a deep learning model that integrates graph learning with spatial modeling for link prediction, enabling high-precision inference of TRNs from spatial ATAC-seq data. Systematic evaluations show that spaNetReg outperforms the existing methods in accuracy and robustness. In practice, our model identifies spatially regionalspecific regulatory networks and core transcription factors in brain development and glioblastoma (GBM), underscoring its biological significance and translational potential.