Single-Cell Gene Network Models Identify Disease Mechanisms Across Specific Cell Types

Computational frameworks leveraging single-cell RNA sequencing data to construct systemic cell-type-specific gene networks can successfully map gene-cell-disease relationships that remain obscured in organism-level bulk tissue datasets, according to a review by researchers Junha Cha and Insuk Lee published in Genomics & Informatics. The authors synthesize recent advances in single-cell network biology, outlining reference-based and de novo computational methodologies that model functional interaction networks across computationally defined cell populations to advance cell-type-resolved disease genetics.

Gene network models provide a foundational framework for graph theory approaches, facilitating the identification of disease genes, drug targets, and underlying genetic mechanisms across diverse biological processes. While mapping genetic interactions yields system-level perspectives on gene functions that extend beyond isolated pathways, interpreting disease genetics requires precise cellular context. Because datasets composed strictly of organism-level bulk samples cannot distinguish context-dependent interactions across mixed cell populations, single-cell RNA sequencing offers a necessary capability to computationally separate cell states and interrogate the cellular drivers of pathology.

Despite the analytical opportunities presented by single-cell transcriptomics, extracting dependable genetic interactions directly from individual cells involves notable technical obstacles. Single-cell RNA sequencing profiles are inherently sparse, driven by the restricted subset of genes expressed by any given cell and compounded by imperfect transcript-capturing efficiencies across measurement platforms. These technical limitations, alongside spontaneous and spurious expression events, substantially hinder standard linear association metrics, such as Pearson correlation coefficients, and non-linear measures, such as mutual information, which often display limited accuracy across diverse cellular contexts.

The review emphasizes that inferring gene relationships within single-cell contexts remains imperative because genetic interactions are dynamic and context-dependent rather than static across an entire organism. In oropharyngeal cancer, for example, expression of the KLRB1 gene encoding CD161 enhances anti-tumor activity in CD4+ follicular helper cells but acts as an immune suppressor when expressed in CD8+ T cells. Similarly, mutations implicated in neurological disorders manifest effects that depend directly on the cellular state in which they arise, underscoring that gene interactions must be evaluated within specific cell types rather than confounded by compositional variations present in bulk tissue samples.

To overcome these challenges, the authors highlight scHumanNet, a reference-based top-down analytical framework that constructs cell-type-specific gene networks using the SCINET algorithm alongside the established HumanNet interactome scaffold. Within this framework, gene activity scores derived from single-cell expression data are transformed from skewed distributions into standard normally distributed subspaces. For each gene pair documented in HumanNet, the minimum activity score is evaluated against a null normal distribution across cell-type subsamplings, and resulting p-values are combined via Fisher’s method into a meta-p-value to determine whether the interaction is preserved within that cell type at its original HumanNet edge weight.

Assessing the validity of modeled cell-type gene networks remains challenging due to the absence of a universally accepted gold standard for cell-type-specific functional gene sets. Evaluating scHumanNet against prior knowledge-based gene sets from public repositories revealed that ordering genes by degree centrality—calculated as the sum of edge weights—reflected cell-type specificity significantly better than alternative reference-free methods or reference-based models utilizing different interactomes. This finding indicates that the chosen scaffold interactome strongly influences network performance and corroborates that network centrality correlates with central, context-specific biological functions.

Beyond basic network construction, scHumanNet incorporates statistical modules for downstream disease analysis, including the detection of differential hub genes between healthy and diseased states through random edge reshuffling null models. This topological approach successfully prioritized disease-relevant genes that exhibited minimal changes in expression levels, demonstrating utility as a complement to traditional differentially expressed gene analyses. Applied to patient-level signatures, the framework deconvoluted cellular mechanisms tied to immunotherapy responses in breast cancer and pinpointed inhibitory neurons in autism spectrum disorder, where disease-associated genes lost network centrality without exhibiting differential expression.

Complementing top-down methods, the authors also describe a de novo bottom-up strategy designed to capture previously uncharacterized interactions unique to disease-associated cell types. This data-driven strategy pairs single-cell expression preprocessing and imputation with Bayesian statistical filtering based on true-positive to true-negative gene pair ratios from curated functional gene sets, effectively mitigating the false positives typically introduced by count matrix transformations or metacell pooling. Benchmarking confirmed that integrating the top-performing preprocessing pipelines captured functionally enriched gene pairs, yielding over 850,000 newly identified co-expression edges that enhanced the predictive performance of reference interactomes in disease gene discovery.

To address the frequent scarcity of matched control samples in single-cell disease studies and mitigate technical batch variability across datasets, the authors developed HCNetlas, a reference database containing 198 cell-type-specific gene networks covering 61 distinct cell types across 25 human organs. Comparing disease-derived networks against healthy reference networks within HCNetlas effectively prioritized known and novel disease genes across systemic lupus erythematosus, lung cancer, and Alzheimer’s disease. Notably, comparative centrality analyses of systemic lupus erythematosus public datasets using HCNetlas reference controls yielded results highly consistent with analyses using internal healthy controls, validating the use of independent atlas-derived controls.

Looking to future applications, the review suggests that cell-type-specific network topologies could serve as input structures for graph neural networks and enable computational perturbation experiments, such as network-based in silico knock-in and knockout simulations using percolation theory or node-failure analysis. The authors also suggest exploring the resilience of network structures against single-cell batch effects to facilitate cross-study meta-analyses, alongside integrating multi-omics datasets—including chromatin accessibility and cell-type-specific mutations—to refine resolution and further elucidate the systemic gene-cell-disease axis.

Functional interaction networks provide a broader biological representation than conventional transcriptional regulatory models that focus solely on transcription factors and their regulons. While transcription factor models delineate direct regulatory binding, functional interaction networks capture a wider spectrum of biological mechanisms, including molecular pathways, gene clusters, epigenetic regulation, and expression quantitative trait loci. The widespread availability of high-throughput next-generation sequencing data in public repositories has accelerated single-cell transcriptomic analyses at reduced costs. In single-cell RNA sequencing, measuring transcriptional activity across individual cells allows co-expression within computationally selected cell populations to serve as a surrogate for genetic interactions. This computational isolation of specific cell states resolves the compositional confounding historically seen in tissue-level bulk samples, where mixed cell types obscured cell-specific genetic contributions.

Within the reference-based scHumanNet pipeline, network construction relies on a statistical scheme designed to balance cell numbers across diverse cell states. The SCINET algorithm converts skewed single-cell expression values into a standard normally distributed subspace to calculate gene activity scores. To prevent sample size imbalances across cell states from distorting inference, the framework performs iterative subsampling across designated cell types. For each gene pair in the HumanNet reference scaffold, the minimum activity score is evaluated against the null normal distribution across subsamples. These significance values are then aggregated into a meta-p-value using Fisher’s combination method. Candidate pairs meeting statistical criteria enter the network with their original HumanNet weights, enabling comparative topology analyses that evaluate centrality shifts through random edge reshuffling null models across distinct cell states.

For de novo network construction, capturing reliable interactions from noisy single-cell datasets requires rigorous quantitative calibration. In this bottom-up framework, gene-gene edge weights are calculated as independent Pearson correlation coefficients across three preprocessing pipelines. These raw correlation scores are converted into log-likelihood scores, and edges with high scores are integrated using an evidence-combination approach identical to that used in HumanNet. Applying a Bayesian filtering step based on true-positive to true-negative ratios from functional gene sets is critical for suppressing false positives introduced by imputation, metacell pooling, or count matrix transformations. When applied to breast cancer metastasis, this data-driven strategy was validated computationally and experimentally, prioritizing metastasis-associated genes and revealing potential therapeutic targets.

The broader utility of cell-type-specific gene networks is highlighted by ongoing cell atlas initiatives aiming to map a periodic table of cells across healthy organs. While such atlases provide valuable reference benchmarks, direct expression comparisons between diseased tissues and healthy atlases remain severely constrained by technical batch variability. Although algorithms can align cell categories across platforms, methods capable of harmonizing expression levels across independent batches remain limited. Because many clinical single-cell investigations lack matched healthy controls, network topology provides an effective comparative alternative. Evaluating changes in network centrality rather than raw expression values allows researchers to use independent healthy atlas samples from HCNetlas as reliable controls, as demonstrated when atlas comparisons closely replicated findings from internal healthy controls in systemic lupus erythematosus studies.

Looking forward, determining the appropriate cellular resolution for network modeling represents an essential computational question. Grouping transcriptionally similar cells to construct networks from cellular heterogeneity shares conceptual similarities with pseudo-bulk methods. Defining the necessary granularity for interpretable networks will require establishing experimentally validated gold-standard gene associations tailored to specific cell types. Such standards will enable systematic assessments of network stability and accuracy across different resolutions. In addition, investigating topological alterations like gene isolation, where disease-relevant genes lose connectivity without showing expression shifts, offers new avenues for understanding pathogenesis. Integrating cell-type-specific network topologies with multi-omic data, such as chromatin accessibility profiles and cell-type mutations, will further clarify how structural network disruptions drive human disease mechanisms.

Subject of Research: Cell-type-specific gene network inference and disease genetics from single-cell RNA sequencing

Article Title: Single-cell network biology enabling cell-type-resolved disease genetics

Article References: Cha, J., & Lee, I. (2025). Single-cell network biology enabling cell-type-resolved disease genetics. Genomics & Informatics, 23(1), Article 10. https://doi.org/10.1186/s44342-025-00042-7

Image Credits: AI Generated

DOI: 10.1186/s44342-025-00042-7

Keywords: single-cell RNA sequencing, gene network biology, scHumanNet, HCNetlas, disease genetics, network centrality, computational biology, transcriptomics, Single-cell, network, biology, enabling