Foundation Model for Single-Cell Omics
Year: 2024
Type: dissertation
Abstract: Single-cell sequencing has enabled unprecedented characterization of distinct cell types and advanced our understanding of disease pathogenesis. The rapid expansion of single-cell data presents both tremendous opportunities and challenges. These rich datasets can reveal cellular heterogeneity, enable lineage tracking, and elucidate pathogenic mechanisms for developing personalized therapeutics. However, effectively analyzing the increasingly massive data also requires scalable computational approaches. As single-cell datasets grow exponentially, a critical need emerges to develop methodologies that can harness, enhance, and adapt to these burgeoning resources, and learning the representation for single-cell omics is one of the core tasks for understanding the exponentially growing sequencing data.The thesis addresses this need by demonstrating a framework to extract meaningful patterns from the single-cell sequencing data. We develop versatile computational models capable of learning from large-scale data. The key innovation is a pretraining and transfer learning paradigm to construct foundation models tailored to single-cell analysis. We design self-supervised learning tasks to enable scalable training and pre-train the transformer foundation model, scGPT, on diverse unlabeled single-cell data to gain general capabilities. We then demonstrate transfer learning by adapting the model to achieve strong performance on specialized downstream tasks: 1) Multi-batch and multi-omic integration, Cell type annotation, and perturbation prediction leveraging the pre-trained cell embeddings. 2) Analyzing developmental dynamics based on newly proposed adaptive neighborhood learning. 3) Continual pre-training expanding the foundation model to spatial transcriptomics. By developing robust generalizable models that inherit knowledge from massive datasets, this thesis provides a blueprint to propel new biological insights in a data-driven way while scaling computational analysis to the exponentially expanding space of single-cell data.
Source: TSpace (University of Toronto)
Author: Cui, Haotian
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7132874304