Boston University
Identifying cell-type-specific dysregulation and prognostic signatures through deconvolution and robust signature extraction
Abstract
dc:description.abstractRNA-Sequencing is one of the most used techniques in investigating the gene expression profiles of biologically interesting phenotypes. Its unbiased detection, high sensitivity and nucleotide-level resolution make it more appealing than other methods such as microarray and qPCR. However, analyzing large-scale datasets of RNA-Seq such as ones from TCGA and extracting biologically meaningful gene expression profiles associated with phenotypes of interests remains challenging. This is due to the variations introduced by both biological/clinical and technical variances, along the high dimensionality of the data. Widely use methods trying to address these issues such as principal component analysis (PCA), non-negative matrix factorization (NMF), and independent component analysis (ICA), are highly sensitive to parameter selection and often yield unstable or non-reproducible results, which limit their interpretability. To address the challenge of the instability and non-reproducibility, this dissertation first introduces ICARus, a computational pipeline designed to extract robust and reproducible gene expression signatures from transcriptomic data using ICA. Different from conventional approaches that rely on deciding a single parameter for the ICA algorithm, ICARus identifies a range of near-optimal parameters based on variance explained by PCA, then performs iterative ICA across these parameters and evaluates the stability and reproducibility of resulting components. By integrating intra- and inter-parameter consistency, ICARus prioritizes high-confidence gene expression signatures that are reproducible across parameter settings. Applications of ICARus to one RNA-Seq dataset of PBMC from COVID patients and one of primary tumors from lung adenocarcinoma demonstrate its ability to identify biologically meaningful signatures associated with clinical outcomes and temporal dynamics, while minimizing parameter-dependent variability. These results support ICARus as a generalizable framework for robust signature extraction in transcriptome datasets with complex clinical and technical settings. During the development of ICARus, biologically meaningful signatures associated with immune functions were revealed constantly in the analysis of lung adenocarcinoma data. This raises the question of whether these gene signatures come from tumor cells or infiltrating immune cells. Bulk RNA sequencing (RNA-Seq) profiles represent mixtures of malignant and non-malignant cell populations, which confounds the gene expression profiles. To overcome this limitation, this dissertation systematically evaluates and applies one published de-convolution algorithm, BayesPrism, to infer cell type fractions and re-construct cell-type specific gene expression profiles from bulk RNA-Seq datasets of 11 cancer types from TCGA. We have demonstrated BayesPrism predicts cell fraction accurately and retains inter-sample heterogeneity in 11 cancer datasets. Then we performed differential, survival and enrichment analyses on the reconstructed gene expression profiles to show that analyses based on deconvoluted expression profiles uncover both shared and cell type-dependent dysregulated and prognosis-associated gene expressions. Together, this dissertation presents two computational frameworks addressing key challenges in transcriptomic analysis: the extraction of robust and reproducible gene expression signatures and cell type-specific transcriptional profiles from heterogeneous bulk RNA-seq data. ICARus provides a reliable approach for identifying reproducible gene expression signatures across diverse datasets while the deconvolution-based analysis framework demonstrates the analysis of cell type-specific expression profiles of biologically/clinically relevant phenotypes. By independently addressing methodological instability and cellular heterogeneity, these approaches enhance the interpretability of transcriptomic data and provide a foundation for more precise investigation of large RNA-Seq datasets.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Bioinformatics
- Grantor
- Boston University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Li, Zhaorong
- Advisors dc:contributor.advisor
-
- Fuxman Bass, Juan
- Labadorf, Adam T.
- ORCID dc:identifier.orcid
Subjects
dc:subject × 1Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53309