Codependency and mutual exclusivity for gene community detection from sparse single-cell transcriptome data

Abstract

Single-cell RNA-seq (scRNA-seq) can be used to characterize cellular heterogeneity in thousands of cells. The reconstruction of a gene network based on coexpression patterns is a fundamental task in scRNA-seq analyses, and the mutual exclusivity of gene expression can be critical for understanding such heterogeneity. Here, we propose an approach for detecting communities from a genetic network constructed on the basis of coexpression properties. The community-based comparison of multiple coexpression networks enables the identification of functionally related gene clusters that cannot be fully captured through differential gene expression-based analysis. We also developed a novel metric referred to as the exclusively expressed index (EEI) that identifies mutually exclusive gene pairs from sparse scRNA-seq data. EEI quantifies and ranks the exclusive expression levels of all gene pairs from binary expression patterns while maintaining robustness against a low sequencing depth. We applied our methods to glioblastoma scRNA-seq data and found that gene communities were partially conserved after serum stimulation despite a considerable number of differentially expressed genes. We also demonstrate that the identification of mutually exclusive gene sets with EEI can improve the sensitivity of capturing cellular heterogeneity. Our methods complement existing approaches and provide new biological insights, even for a large, sparse dataset, in the single-cell analysis field.

Integrated single-cell transcriptomic and epigenetic study of cell state transition and lineage commitment in embryonic mouse cerebellum

“…S5E). To evaluate codependency and mutual exclusivity for Atoh1 , Ascl1 , and other genes in proliferative NPCs, we calculated the codependency index (CDI) and exclusively expressed index (EEI) ( 29 ). We found that Atoh1 and Ascl1 were expressed in a mutually exclusive manner among proliferative NPCs (fig.…”

“…The proliferative NPCs were digitally isolated from scRNA-seq data, and the coexpressed and mutually exclusive gene pairs were calculated ( 29 ). The gene pairs containing Atoh1 or Ascl1 with high CDI (CDI > 10.0) and EEI (EEI > 3.0) are shown in data S3.…”

Multi-level cellular and functional annotation of single-cell transcriptomes

“…Differential expression analyses were performed using Wilcoxon rank sum (Wilcox) and codependency index (CDI) 61,62 . The Wilcox method was implemented using the wilcoxauc function ( Presto R package, v 1.0.0) 63 .…”

“…CDI is then defined as: We further normalized the CDI score using the CDI score corresponding to the probability of observed a perfect co-dependency for cluster k : where π k,k = P ( c k = 1, c k = 1), under the assumption of independence. Possible values of nCDI range between [0,1], such that nCDI = 1 represents perfect co-dependence between a gene and cluster, and nCDI = 0 represents no co-dependence but is not equivalent to mutual exclusivity which has been formulated elsewhere 62 .…”

“…where 𝜋 𝑘,𝑘 = 𝑃(𝑐 𝑘 = 1, 𝑐 𝑘 = 1), under the assumption of independence. Possible values of 𝑛𝐶𝐷𝐼 range between [0,1], such that 𝑛𝐶𝐷𝐼 = 1 represents perfect co-dependence between a gene and cluster, and 𝑛𝐶𝐷𝐼 = 0 represents no co-dependence but is not equivalent to mutual exclusivity which has been formulated elsewhere 62 .…”