Computational integration and trajectory inference of single cell multi-modal data
Year: 2025
Type: dissertation
Abstract: Single-cell analysis provides a new approach to inspect biological processes at a single-cell resolution. Recently, new sequencing protocols have been developed to simultaneously profile the transcriptome, epigenome, and proteome features in single cells. With these technologies, researchers can interpret two or more biological phenomena such as gene regulation and transcription factor binding events at the same time improving the ability to draw causal relationships between these distinct molecular mechanisms. However, a major challenge in single-cell multimodal analysis is that the feature spaces of distinct modalities are extremely different. The differences in feature sizes, sparsity and distributions pose a significant challenge in utilizing the features to obtain a shared latent space across all modalities for downstream analysis. Although several methods have been proposed to address this issue, these methods either require prior knowledge to set parameters or lack scalability. Moreover, several methods only work for specific modality types or do not allow the interpretation of inferred latent components. Another critical issue when dealing with single-cell multimodal data is to infer trajectories to capture cell lineage development. Many methods have been developed to infer trajectories for single-cell data, but few are designed for single-cell multimodal analysis, which can offer molecular information on gene expression and gene regulation at the same cells. Moreover, current approaches have been only used and applied to simpler cell differentiation trees and are unlikely to scale to large trees. While cellular graphs are widely used as representations of single cell data, most methods only use signals on nodes (cells), but do not consider signals associated to edges (differentiation events) between cells. In this thesis, we propose MOJITOO and PHLOWER, accounting respectively for single-cell multimodal integration and trajectory inference. MOJITOO explores Canonical Correlation Analysis (CCA) for an effective and efficient integration of arbitrary modalities into a common joint embedding. Moreover, MOJITOO has few free parameters and allows interpretation, i.e. associates latent spaces to variables. We performed comprehensive benchmarking on multimodal single cell data, which evaluated the MOJITOO properties in the preservation of information in original modalities and in downstream tasks such as distance estimation and clustering. This indicates that MOJITOO performs quite well comparing to competing approaches and has overall lower computational requirements among evaluated methods. In a case study with blood cells, we demonstrate that the latent space obtained by MOJITOO can capture major blood cell types and demonstrate the relation of latent dimensions to known molecular markers. Overall, these results demonstrate that MOJITOO is a powerful computational approach in biological studies for single-cell multimodal integration analysis. PHLOWER is a novel trajectory inference model for multimodal single cell data. It uses simplicial complex and Hodge Laplacian (HL) decomposition to find embedding at edge/trajectories spaces. A comprehensive benchmarking on complex cell differentiation trees indicates that PHLOWER has the best properties in the recovery of tree topologies and associating cells to the trees than state-of-the-art methods. Moreover, we explore the power of multimodal analysis using PHLOWER on large-scale single-cell multimodal data with transcriptome and epigenome in a kidney organoid time course. PHLOWER infers trajectories related to major kidney cells detected in the organoids. Moreover, it detects transcription factors that regulate lineage gene expression. Our analyses shed novel light on mechanisms of kidney lineage development. Altogether, these results demonstrate that PHLOWER is a powerful computational approach in biological studies for single-cell multimodal trajectory inference analysis.
Source: RWTH Publications (RWTH Aachen)
Author: Cheng, Mingbo
Institution: RWTH Aachen University
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7105749716
Cites:
Cited by:
Related to: