Electronic Health Record-Based Computational Drug Repurposing for Head and Neck Cancers with Validation Using Genomic Data and Large Language Models

Head and neck cancers (HNC) represent the seventh most common cancer worldwide, yet the development of novel therapeutics remains slow because traditional drug discovery pipelines are costly, time‑intensive, and prone to failure. Drug repurposing, the identification of new therapeutic indications for existing FDA‑approved medications, offers a pragmatic alternative, particularly when supported by real‑world clinical evidence. This dissertation presents a multi‑layered computational framework that integrates electronic health records (EHR), genomic data, and large‑scale biomedical literature to systematically identify and validate drug repurposing candidates in HNC. Chapter 2 introduces an EHR‑driven analytical pipeline using longitudinal clinical data from the North Carolina Translational and Clinical Sciences (NCTraCS) Institute. Demographic variables, diagnoses, procedural codes, and medication histories are analyzed using machine‑learning classifiers and survival models to identify drugs associated with reduced metastatic risk among HNC patients. Feature importance analyses, Cox analysis, log‑odds ratios, and confounder assessments reveal clinically meaningful, outcome‑driven signals, enabling the prioritization of medications with consistent protective associations in real‑world HNC populations. Chapter 3 extends these findings by incorporating genomic evidence from The Cancer Genome Atlas (TCGA) to provide molecular context and mechanistic validation. Copy number variation and somatic mutation analyses identify genes significantly altered across HPV‑positive and HPV‑negative HNC subtypes. Network‑based protein–protein interaction expansion further captures relevant biological neighborhoods. Mapping these genes and their interaction partners to drug-gene interaction databases enables the alignment of EHR‑derived medication signals with established and emerging molecular drivers of HNC. Chapter 4 integrates large‑scale literature mining using PubMed abstracts and a large language model to extract peer‑reviewed evidence linking candidate genes to HNC biology. Intersecting literature‑supported gene associations with TCGA‑derived risk genes further strengthens mechanistic confidence in prioritized targets and the drugs that modulate them. This triangulation across EHR outcomes, genomic drivers, and literature evidence yields a refined set of high‑confidence repurposing candidates. Across all analyses, several medications consistently emerge, including levothyroxine, melatonin, methylprednisolone, acetaminophen, aspirin, heparin, and dexamethasone. Collectively, this dissertation establishes a scalable, data‑driven framework for computational drug repurposing in head and neck cancer, providing a translational bridge from retrospective clinical observation to mechanistically informed therapeutic hypothesis generation.