Xin He
BME PhD Defense Presentation

Date: 2026-11-18
Time: 2:00 PM–4:00 PM
Location / Meeting Link: In person: 3115 Whitaker Building, 313 Ferst Dr NW, Atlanta, GA 30332 Zoom: https://gatech.zoom.us/j/99837272592?pwd=wXmWGKOYUH5EzZKF4k8xvZSlaJhIoY.1

Committee Members:
Cassie S. Mitchell, PhD (Advisor); Elizabeth M. Cherry, PhD; Ying Zhang, PhD; Duen Horng (Polo) Chau, PhD; May D. Wang, PhD


Title: Transforming Biomedical Literature Into Scientific Discovery: Unifying Graph-Structured Evidence And Large Language Models

Abstract:
Biomedical research increasingly depends on literature-based discovery, which finds hypotheses that the literature implies but no single paper states, yet it faces three challenges: synthesis of scattered graph evidence, relation extraction that ignores task direction, and the credibility of claims over time. This dissertation develops a unified computational framework that transforms large-scale biomedical literature into traceable, directional, and current hypotheses by integrating a PubMed-derived knowledge graph with large language models (LLMs). The framework progresses from synthesizing structured graph evidence into grounded mechanism hypotheses within an interactive platform, to aligning relation extraction with the therapeutic direction that drug repurposing needs, and ultimately to assessing evidence credibility through time-aware detection of medical reversals. The key contribution is a principle shared by all three studies: the literature behind a knowledge graph must be organized, read rather than merely counted, and read over time. By bridging discovery, explanation, and reassessment on one graph, this work advances literature-based discovery from association-driven knowledge mining toward evidence-grounded scientific inference, with applications in mechanistic interpretation, drug repurposing, and the detection of overturned treatments.