Aims: Single-cell RNA-sequencing (RNA-seq) enables high-resolution gene regulatory network (GRN) analysis in specific cell types, but data sparsity, noise, and complex regulatory relationships remain major challenges. Existing methods often focus on pairwise gene associations and insufficiently capture global network topology. This study aims to develop a deep graph framework for single-cell GRN inference by integrating global regulatory structure with biologically informed distributional regularization.

Methods: We propose ZINB-GRAN, a graph adversarial framework for single-cell GRN inference. It constructs a weighted gene co-expression matrix as a prior regulatory graph and reformulates GRN inference as a link prediction task. A graph convolutional encoder learns latent gene representations, while a decoder reconstructs network topology. To enhance biological consistency, a multilayer perceptron-based discriminator aligns encoder-derived representations with a continuous zero-inflated negative binomial (ZINB)-derived prior generated through ZINB sampling, logarithmic transformation, normalization, and Gaussian perturbation.

Results: ZINB-GRAN jointly optimizes network reconstruction and latent distribution alignment using mask-based supervised classification and adversarial losses. This strategy improves regulatory structure discrimination and robustness in sparse single-cell data. Benchmarking on simulated and real datasets shows that ZINB-GRAN outperforms most existing GRN inference methods and identifies cell type-specific GRNs and key regulatory factors in human peripheral blood mononuclear cells (PBMCs) and triple-negative breast cancer.

Conclusion: ZINB-GRAN integrates global network topology, graph convolutional representation learning, and continuous ZINB-derived prior regularization for single-cell GRN inference. By aligning latent representations with a biologically motivated prior, it improves the robustness and interpretability of GRN reconstruction and provides a useful tool for cell-specific network inference, key regulator identification, and biomarker discovery.