Title: Leveraging Prior Knowledge in Multi-Omic Transformers to Infer Single-Cell Gene Regulatory Networks

Abstract:

Gene regulatory networks (GRNs), which map the interactions between transcription factors and their target genes, are fundamental to deciphering cell-state dynamics and disease mechanisms. While reconstructing these networks from paired single-cell transcriptomic and chromatin accessibility data holds significant promise, the task is fraught with difficulties. Specifically, single-cell ATAC-seq (scATAC) data suffers from extreme sparsity, and existing approaches often depend on rigid peak-to-gene mappings alongside limited supervisory signals.

To address these challenges, we introduce EpiAwareNet, a novel framework that utilizes lightweight biological priors within a multi-omic Transformer architecture to infer GRNs from paired single-cell datasets. The methodology operates in two distinct stages. First, EpiAwareNet employs a cross-attention module to learn joint gene-peak representations. This allows for the data-driven, gene-specific aggregation of accessibility signals, moving away from fixed, hard-coded peak-to-gene assignments. Second, the model integrates a GRN prior derived from bulk data, treating it as noisy positive edges to offer weak supervision in the absence of abundant labels. This approach refines regulatory scores while maintaining robustness against potential errors in the prior data.

Our experimental results demonstrate that EpiAwareNet outperforms representative single- and multi-omic baseline methods in GRN reconstruction. Furthermore, the inferred networks exhibit enhanced biological plausibility, evidenced by a superior recovery of established regulatory interactions. These findings suggest that incorporating lightweight biological priors from bulk data, when paired with adaptive cross-modal representation learning, can significantly improve single-cell GRN inference. The source code and data associated with this study will be accessible at https://github.com/tianyang-x/EpiAwareNet_pub.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC