Literature Review: Construction of cell-specific networks from scRNA-seq data

Single-cell RNA sequencing (scRNA-seq) is a powerful technique that allows researchers to analyze gene expression at the individual cell level. It reveals cellular diversity and identifies rare cell types that would otherwise be hidden in bulk RNA sequencing. The workflow involves isolating single cells, capturing their RNA, and using high-throughput sequencing to profile their transcriptomes. This technology has wide-ranging applications in developmental biology, cancer research, immunology, and more.

However, scRNA-seq data often suffer from high technical noise and low sequencing coverage compared to bulk RNA-seq, making it challenging to accurately measure gene expression levels. These technical limitations can lead to false negatives and difficulties in detecting low-abundance transcripts. To address these issues, specialized computational methods are required to normalize the data, reduce noise, and extract meaningful biological insights.

This review aims to examine how Cell-Specific Networks (CSN) and Conditional Cell-Specific Networks (c-CSN) help overcome the challenges of high technical noise and low coverage in scRNA-seq data. It highlights how these network-based approaches improve the accuracy of gene expression analysis and enable deeper biological insights.

Traditional methods for scRNA-seq data analysis often rely on gene expression-based techniques to reduce complexity and visualize high-dimensional data. Principal Component Analysis (PCA) is widely used to identify major sources of variation by projecting data onto principal components. Non-negative Matrix Factorization (NMF) decomposes the data into biologically interpretable features, helping to uncover underlying cellular states. t-distributed Stochastic Neighbor Embedding (t-SNE) is popular for visualizing cell clusters in low-dimensional space, preserving local relationships between cells. However, these methods may struggle with scalability, interpretability, or sensitivity to noise in large and complex scRNA-seq datasets.

Clustering methods play a crucial role in scRNA-seq data analysis by grouping cells with similar gene expression profiles into distinct cell types or states. Hierarchical clustering organizes cells in a tree-like structure, allowing visualization of relationships at multiple resolution levels. K-means clustering partitions cells into a predefined number of groups based on expression similarity, though it may be sensitive to initial parameter settings. More advanced techniques like SNN-Cliq (Shared Nearest Neighbor clustering) leverage graph-based approaches to improve robustness and accuracy in identifying complex cellular subpopulations. These methods collectively enable researchers to uncover the underlying cellular heterogeneity present in single-cell datasets.

Traditional scRNA-seq analysis methods often fail to capture cell type-specific gene interactions, limiting the understanding of unique regulatory mechanisms within distinct cell populations. These approaches typically focus on individual gene expression levels rather than the complex interactions between genes. As a result, they lack a network-level perspective, which is essential for uncovering how genes work together in specific cellular contexts. This limitation highlights the need for more advanced models that integrate gene regulatory networks with single-cell data.

Cell-Specific Networks (CSN) provide a framework for transforming single-cell gene expression data into gene association networks tailored to individual cells, capturing the unique regulatory interactions within each cell. In a CSN, each cell is represented as a network where nodes correspond to genes and edges denote potential functional relationships or interactions between gene pairs, often inferred from expression correlations. Mathematically, given cells and genes, for each cell , an edge between two genes and is defined as , indicating the absence or presence of an interaction based on context-specific gene activity.

In CSN construction, represents the total number of cells in the single-cell RNA sequencing (scRNA-seq) dataset. This parameter is fundamental for defining the statistical framework that infers gene-gene dependencies in individual cells. For each cell , "neighborhoods" around the expression values of genes and are defined using counts and , which are typically set as proportional subsets of (e.g., in the referenced approach). These neighborhoods quantify how many cells exhibit expression levels close to those in cell , enabling localized dependency analysis.

The core statistic measures deviations from statistical independence between genes and in cell , normalizing co-expression counts () by the total number of cells (). The standard deviation , used to normalize into a test statistic, incorporates to reflect how sample size influences the reliability of dependency estimates. Asymptotically, follows a normal distribution parameterized by , allowing hypothesis testing (: independence; : dependency) to determine if an edge exists in the CSN for cell .

The original Cell-Specific Network (CSN) method faces a key limitation: it overestimates gene-gene associations by including both direct and indirect interactions, leading to dense, inaccurate networks that may misrepresent true regulatory relationships.

To address this, conditional Cell-Specific Networks (c-CSN) introduce conditional independence testing, which filters out indirect associations by evaluating gene pairs while conditioning on other genes (e.g., key regulatory hubs). This allows c-CSN to identify direct associations uniquely relevant to each cell, producing sparser, more biologically meaningful networks.

Conditional independence in c-CSN refers to evaluating whether two genes (, ) are independent when conditioning on a third gene (), allowing the filtration of indirect associations mediated by . By testing if the dependence between and vanishes when is known, c-CSN distinguishes direct interactions (remaining dependent) from indirect ones (becoming independent), thereby eliminating spurious connections. This process constructs sparser gene association networks that exclusively capture direct regulatory relationships, reducing overestimation inherent in traditional CSN. The resulting networks are more biologically accurate, reflecting true molecular interactions within each cell, and enable precise downstream analyses like cell clustering and identification of key regulatory modules by focusing on functionally relevant direct edges.

In cCSN (conditional cell-specific network) construction, the framework extends CSN by incorporating conditional independence to filter indirect gene-gene associations. Here, remains the total number of cells in the scRNA-seq dataset, but analyses now focus on three-dimensional neighborhoods involving a pair of genes () and a conditional gene () for each cell .

For each cell , "neighborhoods" are defined around the expression values of , , and :

These counts quantify co-expression patterns conditioned on , enabling the detection of direct associations between and that are not mediated by .

cCSN’s conditional analysis introduces higher computational complexity than CSN, particularly when testing multiple conditional genes or processing large scRNA-seq datasets with thousands of genes and cells.
Developing parallel computing frameworks or optimizing algorithms to scale cCSN for large-scale datasets, reducing runtime while maintaining statistical rigor in conditional dependency inference.

Current c-CSN frameworks infer statistical dependencies rather than causal relationships, limiting their capacity to determine directional interactions (e.g., vs. ) or disentangle direct regulatory effects from unmeasured confounders. This stems from the absence of causal assumptions or temporal ordering in conditional independence tests.

To address this, we propose integrating structure causal models (SCMs) with the cell-specific neighborhood definition from CSN (i.e., the parameter , representing local cell counts in expression neighborhoods). Specifically, we aim to adapt the Peter Clark Algorithm (a causal discovery method) by modifying its conditional independence tests to leverage CSN’s neighborhood-based probability estimates (e.g., ) for (single-cell) data. By embedding cell-specific dependency metrics into causal inference frameworks, this approach could deduce directional, cell-specific causal networks (e.g., directed acyclic graphs, DAGs), enabling the identification of regulatory hierarchies (e.g., transcription factors → target genes) while accounting for cellular heterogeneity. This would bridge statistical association with causal biology, enhancing the interpretability of CCSNs in regulatory pathway reconstruction.