BRIDGE: Biological Evidence Refinement and Heterogeneous Dynamic Gating for Gene Regulatory Networks

Ziyang Dong, Shanwen Tan, Hengchuang Yin, Wei Liu, Yifan Wang, Siyu Yi, Jiancheng Lv, Wei JuarXiv 2026 · arXiv

TL;DR

BRIDGE is a novel graph neural network framework that significantly improves the inference of Gene Regulatory Networks (GRNs) from single-cell RNA sequencing data. By replacing blind, random graph data augmentation with biologically guided edge refinement and introducing dynamic gating to manage communication between gene and cell nodes, BRIDGE achieves state-of-the-art performance and exceptional robustness against noise.

What Problem Does This Paper Address?

Uncovering how transcription factors (TFs) regulate target genes is essential for understanding cellular behavior and disease mechanisms. Single-cell RNA sequencing (scRNA-seq) provides a high-resolution view of gene expression across individual cells, making it a prime resource for mapping these Gene Regulatory Networks (GRNs).

However, scRNA-seq data is notoriously noisy and sparse due to "dropout" events where active genes go undetected. While Graph Neural Networks (GNNs) have recently been applied to predict missing regulatory links, they suffer from two major flaws:

  1. Biologically blind augmentation: Standard GNN contrastive learning relies on randomly dropping edges to create alternative graph views, which often destroys critical biological structures.
  2. Uncontrolled information flow: Existing methods struggle to properly regulate how information propagates between distinct domains (genes vs. cells), leading to noise propagation and blurred cell-state-specific regulatory programs.

Methods

The authors introduce BRIDGE (Biological Evidence Refinement and Heterogeneous Dynamic Gating for Gene Regulatory Networks), a three-stage framework designed to tackle these limitations:

Overall framework of BRIDGE
  1. Biological Evidence-based Edge Refinement (BER): Instead of randomly deleting edges for contrastive learning, BRIDGE calculates a "co-activation strength" for gene pairs based on how frequently they are highly expressed together. It then prunes weakly supported edges to construct a denoised, biologically grounded graph view.
  2. Heterogeneous Dynamic Gated Representation Learning: BRIDGE builds a heterogeneous graph containing gene nodes, cell nodes, and their cross-type relationships. It uses a relation-aware graph attention mechanism (GAT) coupled with a dynamic gating mechanism. This gating mechanism acts as an adaptive filter, controlling exactly how much cross-modal information passes between genes and cells to prevent noise propagation.
  3. Dual-Space Neighborhood Contrastive Learning & Decoding: The model aligns representations across the original and refined graph views in both the gene and cell spaces. Finally, a cell-conditioned decoder uses the dynamic cellular context to score the likelihood of a regulatory interaction between a TF and a target gene.
Comparison of edge-removal augmentations: Random vs. Expression-Aware

Datasets

The framework was evaluated using benchmark scRNA-seq datasets from the BEELINE evaluation suite, spanning:

  • Seven cell types: Five mouse cell types (mESC, mHSC-E, mHSC-GM, mHSC-L, mDC) and two human cell types (hESC, hHEP).
  • Three ground-truth network types: Cell-type-specific ChIP-seq networks (high quality), Non-specific ChIP-seq networks (to test generalization), and the STRING functional interaction network (weak supervision).

Key Insights and Ideas

  • Domain-Aware Pruning Beats Random Perturbation: Random edge dropping treats all connections equally, frequently destroying highly informative network hubs. Pruning based on single-cell co-activation ensures that the underlying backbone of the regulatory structure remains intact during self-supervised learning.
  • Gating as a Noise Insulator: Genes and cells represent fundamentally different biological entities. Allowing unconstrained message passing between them causes cell-level technical noise to corrupt gene embeddings. Dynamic gating effectively insulates the two spaces, letting through only the most robust signals.

Key Results

  • State-of-the-Art Performance: BRIDGE outperformed seven prominent supervised GRN inference methods (including GCLink, HGATLink, and GENELink) in both AUROC and AUPRC across nearly all experimental settings.
  • 5% AUPRC Boost: On cell-type-specific networks, BRIDGE achieved an average 5% improvement in AUPRC over the runner-up baseline, GCLink.
  • Excellent Few-Shot Transfer: In cross-cell-type transfer tests (training on mouse embryonic stem cells and testing on other cell types with only 5% tuning data), BRIDGE consistently maintained superior predictive accuracy.
  • Biological Validation: In a case study on human embryonic stem cells (hESC), BRIDGE predicted novel interactions. Out of the top 10 highest-confidence novel predictions, 9 were independently validated by the ChIPBase database, demonstrating strong real-world utility.

Limitations and Future Work

The authors note that the inferred regulatory relationships are ultimately based on statistical associations rather than direct experimental evidence of causal regulation. Consequently, top predictions still require downstream experimental validation.

For future work, the authors plan to integrate causal representation learning and structural causal models (SCMs) directly into the GNN architecture to transition from correlation-based link prediction to true causal reasoning in gene regulation.

Summary updated

Abstract

Motivation: Gene regulatory network inference from single-cell RNA sequencing (scRNA-seq) data is important for uncovering cell-state-specific transcriptional programs. However, scRNA-seq measurements are sparse and noisy, and experimentally validated TF-target interactions remain limited, making reliable inference challenging. Although graph neural networks have advanced GRN prediction, existing methods often rely on biologically unconstrained graph augmentation, such as random edge perturbation, and insufficiently control information transfer between genes and cells. These limitations may distort regulatory structures and weaken robustness under noisy and weakly supervised settings. Results: To address these issues, we propose an innovative framework named Biological Evidence Refinement and Heterogeneous Dynamic Gating for Gene Regulatory Networks (BRIDGE). BRIDGE extracts gene and cell representations from the expression matrix and its matrix dual, and performs contrastive learning in the gene space and cell space between self and neighbors across the co-expression-refined regulatory view and the original graph. It then applies heterogeneous gated encoding to adaptively regulate information transfer between genes and cells, enabling robust transcription factor-to-target gene prediction. Experiments on benchmark datasets spanning three network types and seven cell types show that BRIDGE achieves state-of-the-art AUROC and AUPRC in most settings. In particular, on Specific networks, BRIDGE improves average AUPRC by 5% over the second-best baseline, GCLink. In cross-cell-type few-shot transfer, BRIDGE consistently outperforms GCLink and GENELink across all six target cell types. A case study on hESC further supports the biological relevance of the predictions, with 9 of the top 10 and 46 of the top 100 novel TF-target interactions validated by ChIPBase.
BibTeX
@misc{dong2026bridge,
  author = {Ziyang Dong and Shanwen Tan and Hengchuang Yin and Wei Liu and Yifan Wang and Siyu Yi and Jiancheng Lv and Wei Ju},
  title = {{BRIDGE: Biological Evidence Refinement and Heterogeneous Dynamic Gating for Gene Regulatory Networks}},
  year = {2026},
  eprint = {2606.14734},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2606.14734},
}

Loading related content...

Figure