Date of Award

2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Program

Biomedical Sciences

Track

Microbiology, Immunology, and Biochemistry

Research Advisor

Weikuan Gu

Committee

Hongsik Cho; Jiafu Ji; Kai Li; Yan Jiao

Keywords

AI, cancer

Abstract

Colorectal cancer (CRC) is a leading cause of cancer-related mortality worldwide, with early diagnosis being critical for improving patient outcomes. This study addresses two key aspects of CRC research: understanding its etiology and prognostic biomarkers and exploring the role of artificial intelligence (AI) in generating hypotheses for early diagnosis. In the first project, we analyzed prognostic biomarkers in CRC, revealing a significant negative correlation between TP53 and CD56 mutations, which have opposing effects on patient survival. TP53, a tumor suppressor gene, is frequently mutated in CRC, while CD56, a glycoprotein involved in cell adhesion, may act as a cancer enhancer. Immunohistochemical (IHC) analysis of 2,923 CRC cases highlighted the potential of these biomarkers in predicting treatment outcomes. In the second project, we evaluated ChatGPT 4.0's capability to generate innovative hypotheses for overcoming challenges in early CRC diagnosis. ChatGPT produced 65 hypotheses across three areas: improving screening accuracy, addressing technological limitations, and identifying reliable biomarkers. While ChatGPT rated 25 hypotheses as "excellent," human evaluators rated only five as highly novel and feasible, emphasizing the need for human oversight in assessing practicality and clinical relevance. Experimental plans for selected hypotheses were developed, with one rated as "excellent" and others as "good" or "moderate." In the third project, we further examined ChatGPT’s ability to generate hypotheses, focusing on drug resistance in CRC treatment. ChatGPT was tasked with generating hypotheses to address five major mechanisms of drug resistance, including genetic mutations, drug transport proteins, alterations in signaling pathways, tumor-associated macrophages (TAMs), and tumor heterogeneity. 68 hypotheses were generated, with 9 rated as novel, 21 as relatively novel, and 38 as not novel. ChatGPT provided experimental designs for each hypothesis, but human evaluators found that AI-generated hypotheses often lacked consideration of key biological constraints and feasibility challenges, reinforcing the necessity of human validation. Together, these studies highlight the importance of biomarker research in understanding CRC progression and the potential of AI as a complementary tool in hypothesis generation. While AI demonstrates remarkable creativity and novelty, human expertise remains essential for evaluating feasibility and translating hypotheses into clinically actionable solutions. This research underscores the need for further validation and collaboration between AI and medical researchers to advance early CRC diagnosis and improve patient outcomes.

ORCID

https://orcid.org/0009-0002-0521-3913

DOI

10.21007/etd.cghs.2025.0694

Download

Share

COinS