scGIR: deciphering cellular heterogeneity via gene ranking in single-cell weighted gene correlation networks

Fei Xu, Huan Hu, Hai Lin, Jun Lu, Feng Cheng, Jiqian Zhang, Xiang Li, Jianwei ShuaiBriefings in Bioinformatics 2024 · DOI

TL;DR

This paper introduces scGIR, a novel algorithm that leverages single-cell gene correlation networks weighted by gene expression to better identify cellular heterogeneity. By applying a PageRank-like algorithm to these networks, scGIR effectively overcomes the noise and sparsity inherent in scRNA-seq data, leading to improved cell type identification and developmental trajectory inference.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) has revolutionized our ability to study cellular diversity within tissues and organisms. However, scRNA-seq data is notoriously noisy and sparse due to technical limitations like "dropout events" (where a gene is detected in some cells but not others, even if it's expressed). This noise makes it challenging to accurately identify distinct cell types, understand their relationships, and infer how cells develop over time. Existing methods often struggle to balance the information from gene expression levels with the underlying biological relationships between genes, leading to potential loss of crucial biological insights.

Methods

The core of scGIR lies in its innovative approach to constructing and analyzing gene networks at the single-cell level. It comprises three main modules:

  1. Gene Correlation Network Construction: This module builds a network where nodes represent genes and edges represent correlations between them. Crucially, it uses a statistical independence measure to determine these correlations, aiming for robustness against noise.
  2. Gene Expression Weighted Module: This is a key innovation. Instead of just having a binary connection (correlated or not), the edges in the gene correlation network are weighted by the actual gene expression levels. This means stronger correlations between highly expressed genes contribute more to the network's structure.
  3. Importance Ranking Matrix Construction: scGIR then applies a modified PageRank algorithm (a random walk model) to this weighted network. This algorithm, inspired by how search engines rank web pages, assigns an "importance score" to each gene within each individual cell. The output is a single-cell gene importance matrix (GIM) with the same dimensions as the original gene expression matrix, but representing gene importance rather than raw expression.

Datasets

The authors evaluated scGIR on nine publicly available scRNA-seq datasets, covering diverse cell types and biological contexts:

  • PBMC4k (peripheral blood mononuclear cells)
  • Mouse bladder cells
  • Li (colorectal tumors and normal mucosa)
  • Chu-time (human embryonic stem cell differentiation)
  • Gokce (various cell types including immune, astrocytes, vascular, neurons)
  • Chu-type (human embryonic stem cell lineage-specific progenitor cells)
  • HND (human neuronal differentiation)
  • Tasic (mouse visual cortex)
  • Trapnell (primary adult myocytes)

These datasets represent a good mix of cell types, numbers of cells, and gene counts, allowing for a comprehensive validation of scGIR's performance.

Key Insights and Ideas

  • Expression matters for network weights: Simply knowing genes are correlated isn't enough; how strongly they are expressed significantly impacts their functional importance and regulatory relationships.
  • PageRank for genes: Adapting a web-ranking algorithm to gene networks provides a powerful way to identify "hub" genes that are central to cellular processes.
  • "Dark" genes revealed: scGIR can identify genes that might not show strong differential expression but are crucial for specific cell states or functions, often missed by traditional methods.
  • Network structure reflects cell state: The complexity and connectivity of gene networks vary significantly between cell types and even subtypes, offering a deeper understanding of cellular identity and function.

Key Results

  • Improved Cell Type Identification: Visualizations (t-SNE) of the GIM generated by scGIR showed clearer separation of cell types compared to using raw gene expression (GEM) or network degree (NDM) alone. For example, scGIR distinctly highlighted the eighth cell type in the PBMC4k dataset.
  • Enhanced Subtype and Trajectory Analysis: In the Chu-type dataset, scGIR improved the identification of subtypes within neural progenitor cells. For the HND neuronal differentiation dataset, scGIR accurately predicted developmental states and trajectories, distinguishing early-stage cells from later ones.
  • Quantitative Performance: Using the Adjusted Rand Index (ARI) across seven clustering methods and nine datasets, scGIR (GIM) consistently outperformed GEM and NDM, demonstrating its superior ability to capture cellular heterogeneity. For instance, in the Chu-time dataset, GIM achieved an ARI of 0.68 with k-means clustering, compared to 0.64 for GEM and 0.53 for NDM.
  • Gene Expression Weighting is Crucial: Comparative analyses showed that scGIR with gene expression weighting significantly improved performance over an unweighted version, highlighting the importance of this feature.

Limitations and Future Work

  • Computational Cost: While scGIR's GIM has the same dimensions as GEM, constructing the weighted gene correlation networks can still be computationally intensive, especially for very large datasets.
  • Dynamic Networks: The authors acknowledge that single-cell weighted gene correlation networks are vast and dynamic, even within the same cell cluster. Future work should focus on uncovering these dynamic characteristics to better understand cell fate decisions.
  • Integration with Multi-modal Data: The paper suggests that future algorithms will need to handle multi-modal single-cell data (e.g., combining transcriptomics with epigenomics or proteomics) and spatial transcriptomics, potentially using multi-layered complex networks and dynamic modeling.
  • Broader Applications: The authors envision scGIR's potential for enhancing drug target prediction and screening efficiency in disease-related scRNA-seq datasets.

Summary updated

Abstract

Single-cell RNA sequencing (scRNA-seq) has emerged as a powerful tool for investigating cellular heterogeneity through high-throughput analysis of individual cells. Nevertheless, challenges arise from prevalent sequencing dropout events and noise effects, impacting subsequent analyses. Here, we introduce a novel algorithm, Single-cell Gene Importance Ranking (scGIR), which utilizes a single-cell gene correlation network to evaluate gene importance. The algorithm transforms single-cell sequencing data into a robust gene correlation network through statistical independence, with correlation edges weighted by gene expression levels. We then constructed a random walk model on the resulting weighted gene correlation network to rank the importance of genes. Our analysis of gene importance using PageRank algorithm across nine authentic scRNA-seq datasets indicates that scGIR can effectively surmount technical noise, enabling the identification of cell types and inference of developmental trajectories. We demonstrated that the edges of gene correlation, weighted by expression, play a critical role in enhancing the algorithm's performance. Our findings emphasize that scGIR outperforms in enhancing the clustering of cell subtypes, reverse identifying differentially expressed marker genes, and uncovering genes with potential differential importance. Overall, we proposed a promising method capable of extracting more information from single-cell RNA sequencing datasets, potentially shedding new lights on cellular processes and disease mechanisms.
BibTeX
@article{xu2024scgir,
  author = {Fei Xu and Huan Hu and Hai Lin and Jun Lu and Feng Cheng and Jiqian Zhang and Xiang Li and Jianwei Shuai},
  title = {{scGIR: deciphering cellular heterogeneity via gene ranking in single-cell weighted gene correlation networks}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbae091},
}
Figure