Author

Zixi XuFollow

Abstract

The elucidation of cell-type–specific signaling networks is central to understanding pancreatic ductal adenocarcinoma (PDAC) and to nominating mechanistically grounded therapeutic targets. We present a Text-to-Target framework that integrates large language models (LLMs) with single-cell omics to couple literature-derived hypotheses to cell-type–resolved expression evidence. Using publicly available datasets, we construct malignant ductal epithelial and lineage-matched acinar meta-cell cohorts from PDAC and perform differential expression analysis to obtain a robust catalogue of disease-associated transcriptional changes. In parallel, an ensemble of LLMs is prompted in a schema-constrained manner to retrieve cell-type–specific targets, pathways, and mechanistic annotations from the biomedical literature. After normalization and quality control, LLM outputs are intersected with PDAC meta-cell DEGs to define an LLM-supported DEG set, which serves as the interface between text priors and omic evidence. We then perform pathway-level integration using over-representation analysis augmented by three LLM-aware scores that quantify pathway recall, expression-weighted activation, and directional concordance, yielding an overall ranking of signaling axes. This integrated analysis recapitulates canonical PDAC modules such as KRAS–MAPK and PI3K–AKT–mTOR, highlights angiogenic and immune checkpoint programs, and elevates replication stress and DNA damage response pathways, including ATR- and PARP-associated circuits, as high-confidence candidates. More broadly, the study demonstrates how this framework can standardize heterogeneous omics and textual knowledge into a unified computational pipeline, enabling reproducible, mechanism-oriented target and pathway discovery in PDAC and, in principle, other complex diseases.

Committee Chair

Fuhai Li

Committee Members

Zachary Abrams, Dan Moran

Degree

Master of Science (MS)

Author's Department

Biomedical Engineering

Author's School

McKelvey School of Engineering

Document Type

Thesis

Date of Award

Winter 12-17-2025

Language

English (en)

Included in

Engineering Commons

Share

COinS