Single-cell Analysis from the perspective of how to Interact, Identify and Integrate cells

Single-cell technologies have emerged as powerful tools to analyze complex tissues at the single-cell resolution, resolving the cellular heterogeneity within a tissue through the discovery of different cell populations. Over the past decade, single-cell technologies have greatly developed allowing the profiling of various molecular features including genomics, transcriptomics and proteomics. These high-throughput technologies produce datasets containing thousands to millions of cells in a single experiment. These large high-dimensional datasets impose several challenges to the data analysis. These challenges can be divided into three categories: interaction, identification and integration. Interaction refers to the visual exploration and interactive analysis of the data, identification refers to the definition of the identity of each single-cell, while integration deals with the combination of different molecular information from different datasets. In this thesis, we introduced several computational methods, addressing these three challenges, to eventually improve the analysis of single-cell data. Regarding the interaction, we focused on developing scalable methods that can analyze datasets having millions of cells and thousands of features within workable time frames. We improved the scalability of both clustering and visualization of single-cell data by summarizing the data using a hierarchical representation. To improve the identification of cells, we make use of the large number of annotated datasets available nowadays, and identify cell populations present in a single-cell dataset using classification methods instead of clustering the data. These classification methods can be trained using the previously annotated datasets. We benchmarked a large number of different classification methods and based on this analysis propose to use simple linear classifiers since they have better performance and scale better to larger datasets. We applied this linear classification on single-cell mass cytometry data to automatically identify cell populations when comparing two cohorts of colorectal cancer patients. To integrate single-cell multi-omics data, we focused on extending the number of measured features to overcome current technological limitations. For single-cell mass cytometry, we integrated different panels measured from the same biological sample, resulting in an extended number of proteins markers per cell. Downstream analysis of this data revealed new cell subpopulations showing a more fine-grained cellular heterogeneity. We also extended spatial single-cell transcriptomic data by integrating it with scRNA-seq data that lacks the spatial localization of the cells. Our proposed integration generates whole transcriptome spatial data, which makes it possible to predict spatial expression patterns of genes (in-silico) that are not originally measured in the spatial data. Taken together, this thesis presents several computational methods that aid and improve single-cell data analysis, increasing our insights in molecular heterogeneity.