Characterization of Antibiotic Resistance Genes using Network-based Approaches

Abstract

Antibiotic resistance is a natural evolutionary response to the selective pressures created by antibiotic use, but its rapid acceleration has become a major public health crisis. Resistance emerges through both vertical inheritance of chromosomal mutations and horizontal transfer of antibiotic resistance genes (ARGs) across diverse bacterial lineages, enabling the spread of drug-resistant infections and undermining effective treatment. Because ARGs circulate across clinical, agricultural, and environmental settings, mitigating resistance requires robust surveillance and mechanistic understanding of how ARGs function, move, and pose risk in real microbial communities. This dissertation develops network-based, multi-scale frameworks for studying ARGs from individual genomes to complex environmental metagenomes. In Chapter 2, we investigate what distinguishes ARGs from other genes within bacterial genomes using protein–protein interaction networks. By applying machine learning to interaction profiles, we identify patterns that differentiate ARGs from non-ARGs and reveal interaction signatures linked to resistance mechanisms and dissemination potential. Chapter 3 extends this analysis to metagenomic settings by extracting ARG genomic neighborhoods from metagenomic assembly graphs, enabling context-aware characterization of ARG mobility and horizontal gene transfer potential within microbial communities. Chapter 4 advances from genomic context to ecological risk by introducing a hazard quantification framework that scores ARGs based on their co-occurrence with mobile genetic elements, virulence factors, pathogens, and other resistance determinants, and applies this framework across diverse environments to study how hazards shift over time in response to external pressures. Finally, Chapter 5 synthesizes these insights into a predictive framework for identifying ARGs directly from metagenomic data. By integrating protein language model embeddings with graph-derived features from gene neighborhood graphs, this context-aware model captures both sequence-level signals and neighborhood structure, improving ARG recovery in complex metagenomic samples. Collectively, this work provides an integrated view of ARGs across biological scales, connecting molecular interaction patterns, genomic neighborhood organization, and environmental hazard to build more accurate and interpretable approaches for resistome profiling and hazard characterization.