Background
Causal inference is crucial in biological research, as it enables the understanding of complex relationships and dynamic processes that drive cellular behavior, development, and disease. Within this context, gene regulatory network (GRN) inference serves as a key approach for understanding the molecular mechanisms underlying cellular function.

While most well-known methods inferring GRNs for single-cell data are limited to construct GRNs at the cell type level, but not at the individual cell level. These methods fail to account for the heterogeneity among individual cells in a population. Other methods like SSN, PSSN have been proposed to construct relevance networks for individual cells. However, these methods primarily focus on correlations between genes, potentially missing causal relationships and more complex regulatory interactions.


Research Progress
In order to address a challenge of inferring causal relationships from single-cell expression profile, Xinzhe Huang et al. at the Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences, have proposed a novel method to construct causal networks for individual cells, referred to as the single cell-specific causal network (SiCNet). By incorporating a statistical concept rooted in causal inference, the research team applied this causal inference framework to construct cell-specific causal networks. This method operates in two steps: (1) constructing a reference causal network from a designated reference dataset, and (2) inferring cell-specific causal networks for individual cells based on the reference network. The reference network serves as the benchmark to capture general causal relationships in gene regulation, while the cell-specific causal network utilizes the gene expression profile of an individual cell in combination with the reference network to infer causal relationships between genes (Fig.1).

By identifying the set of target genes for each regulator, the regulatory activity of each gene can be quantified by counting the number of target genes in the cell-specific network. Based on this, a network outdegree matrix (ODM) was generated for further analysis (Fig. 1d). The ODM retains the same dimensions as the original gene expression matrix (GEM), with the same number of rows and columns, but it represents the higher-order regulatory information of each gene in the network.

The results suggested that the ODM enhances the resolution and clarity of cell type distinctions, offering superior performance in terms of visualizing complex and high-dimensional data compared with the traditional GEM (Fig.2). Additionally, by analyzing cell-specific causal networks, SiCNet can identify key regulators and potential regulatory relationships in the context of cancer, shedding light on cancer progression. Furthermore, by integrating these cell-specific causal networks, SiCNet can represent the states of individual samples and detect critical states by calculating dynamic network biomarkers (DNBs) scores (Fig.3). Moreover, the research team aim to uncover regulatory processes involved in cell reprogramming and dynamic regulatory mechanisms during cellular differentiation, providing insights into factors associated with reprogramming efficiency and hematopoietic specification (Fig.4 and Fig.5).

Overall, SiCNet can reveal cellular heterogeneity and quantify gene regulation with higher throughput and resolution. This approach offers a broader perspective than individual gene expression analysis, and paves a way for deeper biological discoveries and investigations.

Future Prospects
GRN inference has long been a fundamental challenge in biology, with many unresolved issues, such as handling high-dimensional data with insufficient sample sizes, constructing networks from multiple samples, and addressing the complexities of time-series data, including temporal dependencies and feedback mechanisms. The SiCNet method incorporates causal inference while constructing networks for individual cells, offering a solution to some of these challenges. Notably, in addition to modeling at the single-cell level, SiCNet can also be applied to single samples to construct sample-specific networks. Moreover, SiCNet can be extended to spatial transcriptomics data, offering a powerful tool for capturing spatially dependent regulatory information that is crucial for understanding tissue architecture and function.

The complete study is accessible via DOI: 10.34133/research.0743