Logo Logo
Help
Contact
Switch language to German
Computational methods for large-scale single-cell RNA-seq and multimodal data
Computational methods for large-scale single-cell RNA-seq and multimodal data
Emerging single cell genomics technologies such as single cell RNA-seq (scRNA-seq) and single cell ATAC-seq provide new opportunities for discovery of previously unknown cell types, facilitating the study of biological processes such as tumor progression, and delineating molecular mechanism differences between species. Due to the high dimensionality of the data produced by the technologies, computation and mathematics have been the cornerstone in decoding meaningful information from the data. Computational models have been challenged by the exponential growth of the data thanks to the continuing decrease in sequencing costs and growth of large-scale genomic projects such as the Human Cell Atlas. In addition, recent single-cell technologies have enabled us to measure multiple modalities such as transcriptome, protome, and epigenome in the same cell. This requires us to establish new computational methods which can cope with multiple layers of the data. To address these challenges, the main goal of this thesis was to develop computational methods and mathematical models for analyzing large-scale scRNA-seq and multimodal omics data. In particular, I have focused on fundamental single-cell analysis such as clustering and visualization. The most common task in scRNA-seq data analysis is the identification of cell types. Numerous methods have been proposed for this problem with a current focus on methods for the analysis of large scale scRNA-seq data. I developed Specter, a computational method that utilizes recent algorithmic advances in fast spectral clustering and ensemble learning. Specter achieves a substantial improvement in accuracy over existing methods and identifies rare cell types with high sensitivity. Specter allows us to process a dataset comprising 2 million cells in just 26 minutes. Moreover, the analysis of CITE-seq data, that simultaneously provides gene expression and protein levels, showed that Specter is able to incorporate multimodal omics measurements to resolve subtle transcriptomic differences between subpopulations of cells. We have effectively handled big data for clustering analysis using Specter. The question is how to cope with the big data for other downstream analyses such as trajectory inference and data integration. The most simple scheme is to shrink the data by selecting a subset of cells (the sketch) that best represents the full data set. Therefore I developed an algorithm called Sphetcher that makes use of the thresholding technique to efficiently pick representative cells that evenly cover the transcriptomic space occupied by the original data set. I showed that the sketch computed by Sphetcher constitutes a more accurate presentation of the original transcriptomic landscape than existing methods, which leads to a more balanced composition of cell types and a large fraction of rare cell types in the sketch. Sphetcher bridges the gap between the scalability of computational methods and the volume of the data. Moreover, I demonstrated that Sphetcher can incorporate prior information (e.g. cell labels) to inform the inference of the trajectory of human skeletal muscle myoblast differentiation. The biological processes such as development, differentiation, and cell cycle can be monitored by performing single cell sequencing at different time points, each corresponding to a snapshot of the process. A class of computational methods called trajectory inference aims to reconstruct the developmental trajectories from these snapshots. Trajectory inference (TI) methods such as Monocle, can computationally infer a pseudotime variable which serves as a proxy for developmental time. In order to compare two trajectories inferred by TI methods, we need to align the pseudotime between two trajectories. Current methods for aligning trajectories are based on the concept of dynamic time warping, which is limited to simple linear trajectories. Since complex trajectories are common in developmental processes, I adopted arboreal matchings to compare and align complex trajectories with multiple branch points diverting cells into alternative fates. Arboreal matchings were originally proposed in the context of phylogenetic trees and I theoretically linked them to dynamic time warping. A suite of exact and heuristic algorithms for aligning complex trajectories was implemented in a software Trajan. When aligning single-cell trajectories describing human muscle differentiation and myogenic reprogramming, Trajan automatically identifies the core paths from which we are able to reproduce recently reported barriers to reprogramming. In a perturbation experiment, I showed that Trajan correctly maps identical cells in a global view of trajectories, as opposed to a pairwise application of dynamic time warping. Visualization using dimensionality reduction techniques such as t-SNE and UMAP is a fundamental step in the analysis of high-dimensional data. Visualization has played a pivotal role in discovering the dynamic trends in single cell genomics data. I developed j-SNE and j-UMAP as their generalizations to the joint visualization of multimodal omics data, e.g., CITE-seq data. The approach automatically learns the relative importance of each modality in order to obtain a concise representation of the data. When comparing with the conventional approaches, I demonstrated that j-SNE and j-UMAP produce unified embeddings that better agree with known cell types and that harmonize RNA and protein velocity landscapes.
clustering, visualization, t-SNE, UMAP, single-cell RNA-seq, multimodal omics, trajectory alignment, subsampling
Đỗ, Văn Hoàn
2021
English
Universitätsbibliothek der Ludwig-Maximilians-Universität München
Đỗ, Văn Hoàn (2021): Computational methods for large-scale single-cell RNA-seq and multimodal data. Dissertation, LMU München: Faculty of Chemistry and Pharmacy
[img]
Preview
PDF
Do_Van_Hoan.pdf

9MB

Abstract

Emerging single cell genomics technologies such as single cell RNA-seq (scRNA-seq) and single cell ATAC-seq provide new opportunities for discovery of previously unknown cell types, facilitating the study of biological processes such as tumor progression, and delineating molecular mechanism differences between species. Due to the high dimensionality of the data produced by the technologies, computation and mathematics have been the cornerstone in decoding meaningful information from the data. Computational models have been challenged by the exponential growth of the data thanks to the continuing decrease in sequencing costs and growth of large-scale genomic projects such as the Human Cell Atlas. In addition, recent single-cell technologies have enabled us to measure multiple modalities such as transcriptome, protome, and epigenome in the same cell. This requires us to establish new computational methods which can cope with multiple layers of the data. To address these challenges, the main goal of this thesis was to develop computational methods and mathematical models for analyzing large-scale scRNA-seq and multimodal omics data. In particular, I have focused on fundamental single-cell analysis such as clustering and visualization. The most common task in scRNA-seq data analysis is the identification of cell types. Numerous methods have been proposed for this problem with a current focus on methods for the analysis of large scale scRNA-seq data. I developed Specter, a computational method that utilizes recent algorithmic advances in fast spectral clustering and ensemble learning. Specter achieves a substantial improvement in accuracy over existing methods and identifies rare cell types with high sensitivity. Specter allows us to process a dataset comprising 2 million cells in just 26 minutes. Moreover, the analysis of CITE-seq data, that simultaneously provides gene expression and protein levels, showed that Specter is able to incorporate multimodal omics measurements to resolve subtle transcriptomic differences between subpopulations of cells. We have effectively handled big data for clustering analysis using Specter. The question is how to cope with the big data for other downstream analyses such as trajectory inference and data integration. The most simple scheme is to shrink the data by selecting a subset of cells (the sketch) that best represents the full data set. Therefore I developed an algorithm called Sphetcher that makes use of the thresholding technique to efficiently pick representative cells that evenly cover the transcriptomic space occupied by the original data set. I showed that the sketch computed by Sphetcher constitutes a more accurate presentation of the original transcriptomic landscape than existing methods, which leads to a more balanced composition of cell types and a large fraction of rare cell types in the sketch. Sphetcher bridges the gap between the scalability of computational methods and the volume of the data. Moreover, I demonstrated that Sphetcher can incorporate prior information (e.g. cell labels) to inform the inference of the trajectory of human skeletal muscle myoblast differentiation. The biological processes such as development, differentiation, and cell cycle can be monitored by performing single cell sequencing at different time points, each corresponding to a snapshot of the process. A class of computational methods called trajectory inference aims to reconstruct the developmental trajectories from these snapshots. Trajectory inference (TI) methods such as Monocle, can computationally infer a pseudotime variable which serves as a proxy for developmental time. In order to compare two trajectories inferred by TI methods, we need to align the pseudotime between two trajectories. Current methods for aligning trajectories are based on the concept of dynamic time warping, which is limited to simple linear trajectories. Since complex trajectories are common in developmental processes, I adopted arboreal matchings to compare and align complex trajectories with multiple branch points diverting cells into alternative fates. Arboreal matchings were originally proposed in the context of phylogenetic trees and I theoretically linked them to dynamic time warping. A suite of exact and heuristic algorithms for aligning complex trajectories was implemented in a software Trajan. When aligning single-cell trajectories describing human muscle differentiation and myogenic reprogramming, Trajan automatically identifies the core paths from which we are able to reproduce recently reported barriers to reprogramming. In a perturbation experiment, I showed that Trajan correctly maps identical cells in a global view of trajectories, as opposed to a pairwise application of dynamic time warping. Visualization using dimensionality reduction techniques such as t-SNE and UMAP is a fundamental step in the analysis of high-dimensional data. Visualization has played a pivotal role in discovering the dynamic trends in single cell genomics data. I developed j-SNE and j-UMAP as their generalizations to the joint visualization of multimodal omics data, e.g., CITE-seq data. The approach automatically learns the relative importance of each modality in order to obtain a concise representation of the data. When comparing with the conventional approaches, I demonstrated that j-SNE and j-UMAP produce unified embeddings that better agree with known cell types and that harmonize RNA and protein velocity landscapes.