UCell: Robust and scalable single-cell gene signature scoring

1In single-cell RNA-seq analysis, gene signature (or “module”) scoring constitutes a simple yet powerful approach to evaluate the strength of biological signals, typically associated to a specific cell type or biological process, in a transcriptome. Thousands of gene sets have been derived by measuring transcriptional differences between different biological states or cell phenotypes, and are collected in public databases such as MSigDB. More recently, large-scale efforts to construct single-cell atlases, are providing specific gene sets that can be useful to discriminate between cell types. For example, Han et al. have used single-cell RNA sequencing to quantify cell type heterogeneity in different tissues and to define gene signatures for >100 human and murine cell types. Given such a gene set, signature scoring aims at quantifying the activity of the genes in the set, with the goal to characterize cell types, states, active biological processes or responses to environmental cues. The Seurat R package is one of the most comprehensive and widely used frameworks for scRNA-seq data analysis. Seurat provides a computationally efficient gene signature scoring function, named AddModuleScore, originally proposed by Tirosh et al.. However, because genes are binned based on their average expression across the whole dataset for normalization purposes, the method generates inconsistent results for the same cell depending on the composition of the dataset. Inspired by the AUCell algorithm implemented in SCENIC, we propose UCell, a gene signature scoring method based on the Mann-Whitney U statistic. UCell scores depend only on the relative gene expression in individual cells and are therefore not affected by dataset composition. We provide a time- and memory-efficient implementation of the algorithm that can be seamlessly incorporated into Seurat workflows.

2UCell calculates gene signature scores for scRNA-seq data based on the Mann-Whitney U statistic. Given a g × c matrix X of numerical values (e.g. gene expression measurements) for g genes in c cells, we first calculate the matrix of relative ranks R by sorting each column in X; in other words, we calculate a ranked list of genes for each cell in the dataset. Because in scRNA-seq not all molecules in the original sample are observed, transcript counts matrices contain many zeros, resulting in a long tail of bottom-ranking genes. To mitigate this uninformative tail, we set rg,c = rmax + 1 for all rg,c > rmax, with rmax = 1500 by default (matching typical thresholds used for quality control for minimum number of genes detected). To evaluate a gene signature s composed of n genes (s1,…,sn), we calculate a UCell score U’ for each cell j in X with the formula:where U is the Mann-Whitney U statistic calculated by:and R′ is obtained by sub-setting R on the genes in signature s.

3We note that the U statistic is closely related to the area-under-the-curve (AUC) statistic for ROC curves, therefore we expect UCell scores to correlate with methods based on AUC scores such as AUCell. Internally, UCell uses the frank function from the data.table package for efficient ranks computations. Large datasets are automatically split into batches of reduced size, which can be processed serially (minimizing memory usage) or in parallel through the future package (minimizing execution time) depending on the available computational resources.

4UCell is an R package for the evaluation of gene signature enrichment designed for scRNA-seq data. Given a gene expression matrix or Seurat object, and a list of gene sets, UCell calculates signature scores for each cell, for each gene set. In the following illustrative example, we applied UCell to a single-cell multimodal dataset of human blood T cells, which were annotated by the authors using both gene (scRNA-seq) and cell surface marker expression (CITE-seq) (Fig. 1A). Provided a set of T cell subtype-specific genes (Table 1), UCell helps interpreting clusters in terms of signature enrichment in low-dimensional spaces such as the UMAP (Fig. 1B). Importantly, UCell scores are based on the relative ranking of genes for individual cells, therefore they are robust to dataset composition. Evaluating a CD8 T cell signature on the full dataset or on CD8 T cells only, results in identical score distributions for CD8 T cells in the two settings (Fig. 1C). Conversely, AddModuleScore from Seurat normalizes its scores against the average expression of a control set of genes across the whole dataset, and is therefore dependent on dataset composition. CD8 T cells analyzed in isolation or in the context of the full T cell dataset are assigned highly different AddModuleScore scores, with median ~1 in the full dataset and median ~0 for the CD8 T cell subset (Fig. 1D). Another widely-used method for single-cell signature scoring, AUCell, is also based on relative rankings and therefore has the same desirable property as UCell of reporting consistent scores regardless of dataset composition. Compared to AUCell, UCell is about three times faster (Fig. 1E) and uses significantly less memory (Fig. 1F). For example, AUCell requires over 64 GB of RAM to process 100,000 single-cells, while UCell uses only 5.5 GB of peak memory (Fig. 1F), making it suitable even for machines with limited computing power.

5UCell is available as an R package at https://github.com/carmonalab/UCell, and is accompanied by vignettes for signature scoring and for seamless integration with Seurat pipelines. Source code to reproduce the results in this manuscript is available at the following repository: https://gitlab.unil.ch/carmona/UCell_demo.

6This research was supported by the 10.13039/501100001711Swiss National Science Foundation (SNF) Ambizione grant 180010 to SJC.

7Massimo Andreatta: Conceptualization, Methodology, Software, Formal analysis, Visualization, Writing - original draft, Writing - review & editing. Santiago J. Carmona: Conceptualization, Methodology, Software, Formal analysis, Writing - original draft, Writing - review & editing, Funding acquisition.

8The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.