Degree

Doctor of Philosophy (PhD)

Department

Environmental Science

Document Type

Dissertation

Abstract

Antibiotic resistance represents an escalating global health challenge driven by the overuse of antibiotics and the environmental dissemination of antibiotic-resistant bacteria and genes. Rivers act as key conduits and reservoirs for resistance dissemination, receiving continual inputs from wastewater effluents, agricultural runoff, and urban sources. This dissertation integrates a review of ARG detection methods with the development of artificial intelligence-based models for improved identification and classification of resistance genes, followed by an application to the Mississippi River.

Three machine learning frameworks, Trans-ARG, GNN-ARG, and ESM-ARG, were developed using transformer networks, graph neural architectures, and protein language model embeddings to capture structural and functional information within protein sequences. The ESM-ARG model, leveraging embeddings from Meta’s ESM-2, achieved higher predictive accuracy than existing models such as DeepARG, ARGNet, and ARG-SHINE. The framework was applied to metagenomic assemblies from the Mississippi River to assess the riverine resistome. Fewer than 1 percent of open reading frames (ORFs) were classified as ARGs, dominated by beta-lactam, aminoglycoside, multidrug and tetracycline resistance classes. Collectively, this research demonstrates that protein embedding-based deep learning substantially enhances ARG prediction and provides scalable, alignment-free tools for environmental resistome monitoring in complex aquatic ecosystems.

Date

11-3-2025

Committee Chair

Hou, Aixin

Available for download on Monday, November 01, 2032

Share

COinS