Variance-adjusted Mahalanobis (VAM): a fast and accurate method for cell-specific gene set scoring.

Hildreth Robert Frost

doi:10.1093/nar/gkaa582

Abstract

Statistical analysis of single cell RNA-sequencing (scRNA-seq) data is hindered by high levels of technical noise and inflated zero counts. One promising approach for addressing these challenges is gene set testing, or pathway analysis, which can mitigate sparsity and noise, and improve interpretation and power, by aggregating expression data to the pathway level. Unfortunately, methods optimized for bulk transcriptomics perform poorly on scRNA-seq data and progress on single cell-specific techniques has been limited. Importantly, no existing methods support cell-level gene set inference. To address this challenge, we developed a new gene set testing method, Variance-adjusted Mahalanobis (VAM), that integrates with the Seurat framework and can accommodate the technical noise, sparsity and large sample sizes characteristic of scRNA-seq data. The VAM method computes cell-specific pathway scores to transform a cell-by-gene matrix into a cell-by-pathway matrix that can be used for both data visualization and statistical enrichment analysis. Because the distribution of these scores under the null of uncorrelated technical noise has an accurate gamma approximation, both population and cell-level inference is supported. As demonstrated using simulated and real scRNA-seq data, the VAM method provides superior classification accuracy at a lower computation cost relative to existing single sample gene set testing approaches.

Highlights

1.1 Single cell transcriptomicsDespite the diversity of cell types and states present in multicellular tissues, high-throughput genome-wide profiling has, until recently, been limited to assays performed on bulk tissue samples
We focus on single sample gene set testing methods, i.e., those that compute a cell-specific statistic for each analyzed gene set to transform a cell-by-gene scRNA-seq matrix into a sample-by-pathway matrix
The class of single sample gene set testing methods, which transform a cell-by-gene matrix into a cellby-pathway matrix, is particular effective for single cell analyses since it enables the full range of standard downstream processing to be performed on the pathway-level rather than on the gene-level

Summary

Introduction

Despite the diversity of cell types and states present in multicellular tissues, high-throughput genome-wide profiling has, until recently, been limited to assays performed on bulk tissue samples. Provide scientists with a detailed picture of cellular biology Such cell-level genomic resolution is especially important for the study of tissues whose structure and function is defined by complex interactions between multiple distinct cell types that can occupy a range of phenotypic states, e.g., the tumor microenvironment [8, 9], immune cells [10, 11], and the brain [12]. Important scientific questions that can be addressed by single cell transcriptomics include the identification and characterization of the cell types present within a tissue [13, 14], the discovery of novel cell subtypes [15], the analysis of dynamic processes such as differentiation [7], or the cell cycle [16], and the reconstruction of the spatial distribution of cells within a tissue [5]

Methods

Results

Conclusion