Date of Award

12-17-2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Bioinformatics

First Advisor

Mary Yang

Abstract

The tumor microenvironment (TME) comprises diverse cell types and plays crucial roles in cancer development and therapeutic response. Single-cell RNA sequencing (scRNA-seq) data quantifies gene expression at the single-cell level, enabling the dissection of heterogeneous cell populations in various biological contexts and diseases. In this work, we developed integrative single-cell genomics and machine learning approaches to characterize the TME and enhance patient survival prediction. We introduced a convolutional neural network (CNN)-based method, cnnImpute, to address the high proportion of missing data in scRNA-seq datasets. cnnImpute effectively imputes missing values while preserving the integrity of cell clusters, achieving superior performance in various benchmarking experiments. It provides a useful resource for scRNA-seq data analysis. To leverage the wealth of information from studies with large patient cohorts and high-resolution transcriptome profiling, we developed a computational framework that integrates bulk and single-cell RNA sequencing data to study the immune landscape and cellular communication within the TME. We identified patient subgroups with varying levels of immune infiltration in breast cancer. Our analysis revealed significant differences in survival rates across these immune infiltration groups, with patients in different groups also exhibiting distinct genomic mutation patterns. In characterizing cell-cell communication within the TME, we constructed cellular communication networks by integrating ligand-receptor interactions with scRNA-seq expression data. Our results revealed significant variations in communication networks among different immune response groups. This work thoroughly characterizes the composition and dynamic interplay of the TME, uncovering associations between immune infiltration and clinical outcomes. The unique mutations and signaling pathways associated with patient subgroups provide insights into the mechanisms underlying diverse tumor immune infiltration and the formation of an immunosuppressive tumor microenvironment. Finally, we introduced a model incorporating a transformer-based encoder to effectively process categorical clinical data and improve the accuracy of survival prediction. Our work offers robust computational methods to facilitate cancer research, leading to a better understanding of disease mechanisms and optimizing treatment strategies.

Download

Share

COinS