Preview

PDF, English - main document Download (39MB) | Terms of use

Abstract

Single-cell RNA-sequencing profiles the transcriptome of cells from diverse populations. A popular intermediate data format is a large count matrix of genes x cells. This type of data brings several analytical challenges. Here, I present three projects that I worked on during my PhD that address particular aspects of working with such datasets:

- The large number of cells in the count matrix is a challenge for fitting gamma-Poisson generalized linear models with existing tools. I developed a new R package called glmGamPoi to address this gap. I optimized the overdispersion estimation procedure to be quick and robust for datasets with many cells and small counts. I compared the performance against two popular tools (edgeR and DESeq2) and find that my inference is 6x to 13x faster and achieves a higher likelihood for a majority of the genes in four single-cell datasets. - The variance of single-cell RNA-seq counts depends on their mean but many existing statistical tools have optimal performance when the variance is uniform. Accordingly, variance-stabilizing transformations are applied to unlock the large number of methods with such an requirement. I compared four approaches to variance-stabilize the data based on the delta method, model residuals, inferred latent expression state or count factor analysis. I describe the theoretical strength and weaknesses, and compare their empirical performance in a benchmark on simulated and real single-cell data. I find that none of the mathematically more sophisticated transformations consistently outperform the simple log(y/s+1) transformation. - Multi-condition single-cell data offers the opportunity to find differentially expressed genes for individual cell subpopulations. However, the prevalent approach to analyze such data is to start by dividing the cells into discrete populations and then test for differential expression within each group. The results are interpretable but may miss interesting cases by (1) choosing the cluster size too small and lacking power to detect effects or (2) choosing the cluster size too large and obscuring interesting effects apparent on a smaller scale. I developed a new statistical framework for the analysis of multi-condition single-cell data that avoids the premature discretization. The approach performs regression on the latent subspaces occupied by the cells in each condition. The method is implemented as an R package called lemur.