News Release

Graph learning and single-cell genomics may unlock more accurate gene regulatory network inference

ZINB-GRAN integrates global network topology with biologically informed distributional regularization to achieve superior performance in identifying regulatory interactions

Peer-Reviewed Publication

Science Exploration Press

Overview of the ZINB-GRAN: Starting with the count matrix from scRNA-seq data as input, ZINB-GRAN first constructs a WGCN from gene expression data.

image: 

Overview of the ZINB-GRAN: Starting with the count matrix from scRNA-seq data as input, ZINB-GRAN first constructs a WGCN from gene expression data. Based on this WGCN, it builds an initial regulatory graph for the genes. The initial regulatory graph and gene expression profiles are then input into a GAE. The GAE model consists of a GCN and a scoring function: the GCNs serve as the encoder, learning the global regulatory structure and embedding it into gene representations, while the scoring function acts as the decoder, scoring the gene pairs’ representations and reconstructing the GRN. The GAE aligns the latent representation Z with the prior distribution. scRNA-seq: single-cell RNA sequencing; WGCN: weighted gene co-expression network; GAE: graph autoencoder; GCNs: graph convolutional networks; GRN: gene regulatory network.

view more 

Credit: © Jianping Zhao, Junfeng Xia, Chunhou Zheng, et al. This is an Open Access article licensed under a Creative Commons Attribution 4.0 International License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, sharing, adaptation, distribution and reproduction in any medium or format, for any purpose, even commercially, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.

A new graph-based deep learning framework may improve the reconstruction of gene regulatory networks from single-cell RNA sequencing (scRNA-seq) data by integrating global network structure learning with biologically informed statistical modeling, according to a study published in Computational Biomedicine.

The method, ZINB-GRAN, addresses major challenges in single-cell gene regulatory network (GRN) inference, including data sparsity, technical noise, and the difficulty of capturing complex, system-level regulatory relationships between genes.

Gene regulatory networks describe how genes interact to control cellular behavior. Although single-cell RNA sequencing enables high-resolution measurement of gene expression at the cellular level, reconstructing accurate regulatory interactions from such data remains challenging.

Most existing approaches focus on pairwise gene relationships and are limited in their ability to capture the global topology of regulatory networks, reducing their effectiveness in complex biological systems.

To address these limitations, the researchers developed ZINB-GRAN, a graph adversarial learning framework that formulates GRN inference as a link prediction problem on a gene co-expression network.

The framework first constructs a weighted gene co-expression matrix as a prior graph representation. A graph convolutional encoder is then used to learn latent representations of genes, while a decoder reconstructs the regulatory network structure.

To improve biological consistency, the model incorporates a distributional regularization strategy based on a zero-inflated negative binomial (ZINB) prior, which reflects the statistical properties of sparse single-cell gene expression data. This prior is transformed into a continuous form through sampling, normalization, and Gaussian perturbation, and is aligned with learned representations using adversarial training.

The model jointly optimizes network reconstruction and latent representation alignment through supervised classification and adversarial objectives, improving robustness in sparse and noisy datasets.

In benchmarking experiments using simulated and real-world datasets, ZINB-GRAN outperformed existing gene regulatory network inference methods. It showed improved performance in reconstructing regulatory network structures and identifying biologically meaningful gene interactions.

Applications to human peripheral blood mononuclear cells (PBMCs) and triple-negative breast cancer datasets demonstrated the model’s ability to identify cell type-specific regulatory networks and key regulatory factors associated with immune function and cancer-related processes.

The authors suggest that integrating global network topology learning with biologically informed statistical priors can improve both the accuracy and interpretability of gene regulatory network inference, providing a useful tool for studying cellular regulatory mechanisms.


Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.