Uncovering cell identity through differential stability with Cepo

Abstract

We present Cepo, a method to generate cell-type-specific gene statistics of differentially stable genes from single-cell RNA-sequencing (scRNA-seq) data to define cell identity. Cepo outperforms current methods in assigning cell identity and enhances several cell identification applications such as cell-type characterisation, spatial mapping of single cells, and lineage inference of single cells.Defining cell identity is fundamental to understand the cellular heterogeneity in the population and the cell-type-specific response to environmental signals and experimental perturbations. Exploring cell identity has been enabled by rapid technological advances in genome-wide profiling of the molecular content in single cells 1-3 . This comprehensive lens into the molecular properties unique to each cell type allows defining and predicting cell identities in ways that were previously not feasible using data generated by bulk/population analytics technologies. The most widely used method to define genes associated with cell identity is differential expression (DE) 4,5 . Despite the advances in high-throughput scRNA-seq data that would provide unprecedented resolution of cell identity, none of the current approaches 6 has been evaluated systematically for their attribute and fidelity for defining cell identity genes (CIGs) from scRNAseq dataset of millions of cells and consisting of hundreds of cell types.We developed Cepo (refers to "cell" in Korean), a method to retrieve genes defining cell identity from scRNA-seq data. We propose a biologically motivated metric, differential stability (DS), to identify cell-type specific genes on the premise that stable gene expression is a key indicator of cell identity. Our hypothesis implies that genes marking a cell type should be (1) expressed and (2) stable in its expression level relative to other cell types. We translate the criteria into a computational framework where, using pre-defined celltype labels, we compute cell-type-specific statistics to prioritise genes that are DS against other cell types in all cell-type pair comparisons (Fig. 1a, Online Methods: Cepo implementation).We performed a comprehensive benchmarking of Cepo with several differential analysis methods 6-9 using both simulated and experimental scRNA-seq datasets in a range of biological systems that require knowledge of cell identity. We tested the accuracy and efficacy of the method to detect subtle changes in stability against simulated DS genes with varying stability (Supplementary Fig. 1 and Online Methods). We show that Cepo, followed by Voom, can identify simulated DS genes with the highest accuracy relative to other methods (Fig. 1b and Supplementary Fig. 2, 3).

Comprehensive Characterisation of Fetal and Mature Retinal Cell Identity to Assess the Fidelity of Retinal Organoids

“…To resolve genes that underlie retinal cellular identity, we computed a cell-type-specific cell identity score for each gene by dataset and batch using Cepo, a computational method for detecting cell identity genes (Kim et al, 2021). The clustering of samples from across datasets and batches using Pearson’s correlation of Cepo-derived gene statistics show strong grouping by cell type irrespective of the origin of dataset and batch ( Figure 2A ).…”

“…To derive the cell identity gene statistics for the human retina atlas, the count matrix of cell-gene variables were first log-transformed and normalized using the logNormCounts function from the scater package (McCarthy et al, 2017). The transformed and normalized data from each batch were subsequently analyzed using the Cepo package (Kim et al, 2021) for quantifying cell identity gene statistics for each major cell type based on the differential stability metric. For comparison, alternative methods (e.g., limma (Ritchie et al, 2015), voom (Law et al, 2014), edgeR (Robinson et al, 2010)) based on differential expression analysis were also used for calculating cell-type-specific gene statistics.…”

Single-cell atlas of bronchoalveolar lavage from preschool cystic fibrosis reveals new cell phenotypes

“…Broad cell type labels were also manually assigned to each cluster, resulting in a total of 13 broad cell type labels across the three data subsets. Marker genes for the consolidated subpopulations in each data subset were identified using the ‘Cepo’ (24) method. Differences in composition of macrophage subpopulations between the control and CF samples were assessed for statistical significance using the ‘propeller’ test from the speckle (version 0.0.3) package (https://github.com/Oshlack/speckle).…”

Computational approaches for direct cell reprogramming: from the bulk omics era to the single cell era

“…As these methods were mostly based on differential expression methods, single cell data provide the opportunity to use ideas from more recent single cell differential expression methods like MAST [ 36 ] which uses a generalized linear model to address the specific biases seen in single cell data. These data also opens up the opportunity for other analyses that depend on the gene expression distribution like differential variability [ 37 ], differential distribution [ 38 ] and differential stability [ 39 ], which may identify TFs with different properties that could be novel candidate TFs for direct cell reprogramming.…”