Graduation Date

Spring 5-9-2026

Document Type

Thesis

Degree Name

Master of Science (MS)

Programs

Molecular Genetics & Cell Biology

First Advisor

Dr. Shibiao Wan

Abstract

The increasing accessibility of transcriptomic profiling has resulted in a surge of molecular pathway-based cancer subtyping in the last decade. Subtype-specific treatment and risk stratification have been demonstrated to improve survival rates and may allow for a more efficient allocation of resources. Pancreatic cancer (PaC), the third leading cause of cancer-related deaths in the United States, is known for its high intra- and inter-tumor heterogeneity. Pancreatic ductal adenocarcinoma (PDAC), the most common variant of PaC, has a very low five-year survival rate and remains difficult to treat. Conventional wet-lab approaches for PDAC subtyping (microdissection, histopathological studies, etc.) are laborious, costly, and often ineffective. Although the basal and classical subtypes of PDAC are widely accepted, high heterogeneity in clinical response and molecular features still exists within these groups. Further subtyping focused on this molecular landscape may allow for better risk stratification and personalized treatment. To address these concerns, we developed two advanced machine learning approaches (i.e., one supervised learning approach and one unsupervised learning approach) to identify PDAC subtypes based on transcriptomic data. Specifically, for the supervised learning approach, we developed a novel stacking-based learning framework, MetaPaCS, to categorize tumor samples into a previously established four-subtype scheme based on RNA-seq data. MetaPaCS demonstrated significantly better performance over existing machine learning (ML) classifiers. For the unsupervised learning approach, we created PacEnClust, a wMetaC-based framework to explore high-resolution PaC subtypes by integrating eleven different cohorts of PaC patients. PacEnClust was shown to outperform traditional clustering algorithms when finding the same number of clusters and revealed a stable five subtype scheme for PDAC. We expect our proposed approaches will provide robust frameworks for better PaC subtype characterization for accurate downstream risk assessment and personalized treatment design. 

Rights

The author holds the copyright to this work and any reuse or permissions must be obtained from the author directly.

Available for download on Thursday, April 22, 2027

Share

COinS