Year: 2019
Type: dissertation
Abstract: Single-cell RNA-sequencing (scRNA-seq) involves the measurement of gene expression from isolated single cells, with the potential to illuminate cellular heterogeneity within complex tissue samples. However, scRNA-seq data are subject to a large number of technical effects. In this work, we present two approaches that can be applied for identifying technical effects in scRNA-seq data in pursuit of the ultimate aim of removing these effects in downstream analyses. The first project introduces different concepts of stably expressed genes with respect to true biological expression. Different classes of stably expressed genes may capture different technical effects, assisting in the removal of these technical effects and increasing the biological interpretability of later results. We define different notions of stability explicitly and organize gene sets by structure to identify gene sets that are systematically enriched with different classes of stably expressed genes. We find that genes associated with the cytosolic ribosomal structure are enriched with genes that are stably expressed in proportion to the total RNA content of a cell. The cytosolic ribosomal genes can serve as a foundation for a set of negative control genes incorporated into normalization procedures to remove some technical effects associated with cell size. The second project describes a procedure to estimate the efficiency with which genes are captured in scRNA-seq experiments. Unique molecular identifiers (UMIs) tag molecules before amplification and allow for later deduplication of reads originating from the same original molecule. Ordinarily, we are interested in the number of UMIs present to enumerate the gene expression of a cell. We consider the information typically discarded during deduplication by measuring the number of reads captured for each unique molecular identifier (rUMI). Features of reliable detection and genes with those features can be identified from the rUMI distributions and applied to downstream analyses, correcting for additional technical effects. Specifically, we estimate the unobserved proportion of molecules that are not captured (0 rUMI) from the observed distribution of rUMI by extrapolation. We propose a framework for three possible zero-truncated integer-based distributions to model the rUMI distribution. Based on the models, we identify five estimators of the true proportion of reads that are lost during the experiment and result in 0 rUMIs. We derive and simulate asymptotic features of these estimators under different conditions and perturbations motivated by scRNA-seq data. We apply our estimators to a scRNA-seq dataset, noting molecule characteristics like the proportion of GC bases in a transcript, that suggest reliable detection. Together, these projects provide two approaches to identifying technical biases in scRNA-seq data and can be applied to later normalization procedures.
Author: Julie M. Deeke
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W3138702931