Computational methods for single cell omics data analysis

The rapid advancement of single cell technologies always brings new challenges in the development of computational methods to effectively analyse increasingly large-scale and high dimensional single cell datasets. In this work, we explore and address several issues in the analysis of single cell data, with a focus on better understanding of cell types. To evaluate clustering methods on their ability to automatically estimate the number of cell types, we performed a systematic benchmarking study to evaluate 14 methods in four categories of different techniques for estimating the number of cell types in single cell RNA data. We uncover the performance differences between these methods and summarise these multi-faceted evaluations into a recommendation. To improve the clustering of multimodal single cell omics data, we present SnapCCESS, an ensemble deep learning framework for generating multi-view integrated embeddings for consensus clustering of multimodal single cell omics data. We show that by using the snapshot ensemble learning technique, snapCCESS can achieve high performance in cell clustering and improve computational speed. To jointly examine the effect of cell types or other potentially related factors of differential abundance signatures, we present iDAS, an ANOVA-based framework for separating gene signatures that are common to all individuals from those that are specific to a particular cell type or cell state. In the case study, we use iDAS to identify the cell type-specific gene signatures that respond to immunotherapy. Overall, this thesis addresses several questions in the analysis of single cell data, focusing on the estimation and development of methods for cell type identification using single cell RNA-seq and multi-omics data, and the use of cell type information for a better understanding of complex gene signatures.