Date of Award

8-13-2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Information Science

First Advisor

Dongying Li

Abstract

Drug-induced cardiotoxicity (DICT) is a major cause of drug attrition during clinical development and post-marketing withdrawal. Despite rigorous preclinical evaluations, cardiotoxic effects often go undetected until late-stage trials or after regulatory approval, posing significant risks to patient safety and leading to costly drug discontinuations. Accurate prediction of DICT remains a substantial challenge. While in vitro and in vivo models provide valuable mechanistic insights, their application is limited by cost, time, and scalability. In contrast, in silico methods—particularly those powered by artificial intelligence (AI)—offer a promising avenue for early safety assessment. This research advances DICT prediction by integrating chemical structure information, regulatory knowledge, and AI techniques to build a comprehensive and scalable framework for early risk evaluation. To support model development, a large-scale drug reference list —DICTrank—was constructed by systematically annotating 1,318 FDA- approved drugs with four levels of DICT concern, based on curated information from FDA-approved drug labeling documents. This dataset spans a broad range of therapeutic areas and represents the most extensive human expert labeled DICT reference to date. Leveraging DICTrank, a series of machine learning models based on quantitative structure–activity relationships (QSARs) were developed and evaluated based on chronological data splitting to simulate real-world scenarios. Among the tested algorithms, Logistic Regression and XGBoost achieved the best performance. Feature importance analysis revealed that structural and physicochemical properties—such as polarizability and electronegativity—were key contributors to DICT prediction. To further improve predictive accuracy and interpretability, this study introduced a novel framework—quantitative knowledge–activity relationships (QKARs)—which integrates pharmacological context and therapeutic knowledge alongside molecular descriptors. QKAR models consistently outperformed traditional QSARs across both cardiotoxicity and hepatotoxicity prediction tasks. In summary, this work presents a structure and knowledge-enriched, AI-driven framework that advances the state of predictive toxicology. By integrating regulatory data, cheminformatics, and modern machine learning, it offers a practical and interpretable tool for early DICT risk evaluation, ultimately supporting safer and more efficient drug development.

Download

Share

COinS