CCSN: Single Cell RNA Sequencing Data Analysis by Conditional Cell-specific Network

Lin Li, Hao Dai, Zhaoyuan Fang, Luonan ChenbioRxiv 2020 · DOI

TL;DR

This paper introduces a new method called Conditional Cell-Specific Network (CCSN) to build more accurate gene interaction networks from noisy single-cell RNA sequencing (scRNA-seq) data. CCSN filters out indirect gene associations, leading to improved cell type identification and a novel metric, Network Flow Entropy (NFE), for quantifying cell differentiation potency.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular heterogeneity, but the data is inherently noisy and has low coverage. This makes it challenging to accurately infer gene interactions within individual cells. Existing methods, like Cell-Specific Network (CSN), can identify gene associations but tend to overestimate them by including indirect effects. This overestimation can obscure true biological relationships and hinder downstream analyses like cell clustering and understanding developmental processes. The authors aim to address this by developing a method that specifically identifies direct gene-gene associations.

Methods

The core of this work is the Conditional Cell-Specific Network (CCSN) method. It builds upon the concept of statistical independence.

  • Conditional Independence: The fundamental idea is that if two genes (x and y) are conditionally independent given a third gene (z), then any association between x and y that is mediated only through z is considered indirect. CCSN leverages this to filter out these indirect associations.
  • Hypothesis Testing: For each gene pair (x, y) in a cell, CCSN performs a hypothesis test to determine conditional independence given a set of "conditional genes" (z). If they are found to be conditionally independent, the direct association is considered absent.
  • Conditional Gene Selection: To make the process computationally feasible, CCSN doesn't test all other genes as conditional genes. Instead, it uses a strategy based on the original CSN to identify a small set of "important" genes (e.g., hub genes) as conditional genes.
  • Network Flow Entropy (NFE): Based on the constructed CCSN, the authors introduce Network Flow Entropy (NFE) as a measure of a cell's differentiation potency. NFE quantifies the "entropy" of the gene interaction network within a cell, with lower entropy potentially indicating a more differentiated state.

Datasets

The authors used several publicly available scRNA-seq datasets for validation, including:

  • Multiple scRNA-seq datasets with varying numbers of cells (100 to 20,000).
  • The Wang dataset (NPC differentiation).
  • The Yang dataset (mouse hepatoblast differentiation).
  • The TCGA lung adenocarcinoma (LUAD) RNA-seq dataset (for bulk RNA-seq analysis).

Key Insights and Ideas

  • Direct vs. Indirect Associations: The key insight is that by conditioning on other genes, we can disentangle direct gene interactions from indirect ones, leading to a more accurate representation of cellular networks.
  • Improved Cell Type Separation: CCSN-transformed data (CNDM) shows superior performance in cell clustering and visualization compared to raw gene expression data (GEM), suggesting that network-based features are highly informative for distinguishing cell types.
  • NFE as a Potency Metric: Network Flow Entropy (NFE) emerges as a promising new metric for quantifying cell differentiation potency, effectively capturing developmental trajectories.

Key Results

  • Clustering Performance: CCSN-derived data (CNDM) consistently outperformed GEM (gene expression matrix) in clustering accuracy across multiple datasets, as measured by the Adjusted Random Index (ARI). For example, in the Buettner dataset, CNDM achieved an ARI of 0.87 with K-means clustering, compared to 0.29 for GEM.
  • Network Sparsity: CCSN networks are significantly sparser than CSN networks, indicating the successful removal of indirect associations.
  • Differentiation Trajectories: NFE successfully illustrated the gradual decrease in differentiation potency during NPC differentiation, aligning with known biological processes. In the Wang dataset, NFE showed a significant decrease from progenitor cells (Day 0) to mature neurons (Day 30).
  • Bulk RNA-seq Applicability: CCSN was also shown to be effective for analyzing bulk RNA-seq data, successfully distinguishing between normal and tumor tissues in the TCGA LUAD dataset.

Limitations and Future Work

  • Computational Cost: The computational cost of CCSN increases with the number of conditional genes (G). The authors suggest that a parallel computation scheme would be beneficial.
  • Causality: CCSN is designed to identify associations, not causal relationships. The direction of gene associations cannot be determined by this method.
  • Future Research: The authors identify the development of parallel computation schemes and methods for inferring causal gene associations as future research directions.

Summary updated

Abstract

Abstract The rapid advancement of single cell technologies has shed new light on the complex mechanisms of cellular heterogeneity. However, compared with bulk RNA sequencing (RNA-seq), single-cell RNA-seq (scRNA-seq) suffers from higher noise and lower coverage, which brings new computational difficulties. Based on statistical independence, cell-specific network (CSN) is able to quantify the overall associations between genes for each cell, yet suffering from a problem of overestimation related to indirect effects. To overcome this problem, we propose the “conditional cell-specific network” (CCSN) method, which can measure the direct associations between genes by eliminating the indirect associations. CCSN can be used for cell clustering and dimension reduction on a network basis of single cells. Intuitively, each CCSN can be viewed as the transformation from less “reliable” gene expression to more “reliable” gene-gene associations in a cell. Based on CCSN, we further design network flow entropy (NFE) to estimate the differentiation potency of a single cell. A number of scRNA-seq datasets were used to demonstrate the advantages of our approach: (1) one direct association network for one cell; (2) most existing scRNA-seq methods designed for gene expression matrices are also applicable to CCSN-transformed degree matrices; (3) CCSN-based NFE helps resolving the direction of differentiation trajectories by quantifying the potency of each cell. CCSN is publicly available at http://sysbio.sibcb.ac.cn/cb/chenlab/soft/CCSN.zip .
BibTeX
@misc{li2020ccsn,
  author = {Lin Li and Hao Dai and Zhaoyuan Fang and Luonan Chen},
  title = {{CCSN: Single Cell RNA Sequencing Data Analysis by Conditional Cell-specific Network}},
  year = {2020},
  doi = {10.1101/2020.01.25.919829},
  howpublished = {bioRxiv},
}
Figure