Computational Approaches to Support Environmental Surveillance of Antimicrobial Resistance

Abstract

Antimicrobial resistance (AMR) is a major threat to global health. Antibiotic resistance genes (ARGs) are released from wastewater treatment plants and livestock farms, and many are carried on mobile genetic elements (MGEs) that can move between bacteria. AMR in these settings has mostly been studied with culture-based methods, which are slow and capture one organism at a time, or with quantitative PCR, which measures only a few pre-selected genes. Non-target chemical analysis and shotgun metagenomic sequencing measure many more chemicals and genes in each sample but require computational methods for interpretation. This dissertation develops and applies computational approaches to three related problems: predicting the removal of pharmaceuticals and personal care products (PPCPs) during wastewater treatment, characterizing airborne ARGs at livestock farms, and improving the reference database used to identify MGE-associated proteins. Using a two-step machine-learning framework that combined unsupervised clustering and supervised classification, this work predicted the removal patterns of 149 PPCPs across two full-scale wastewater and water-reuse treatment trains from chemical descriptors. Classification accuracy ranged from 42.5% to 65.2%, depending on the facility and clustering approach. PPCP clusters based on measured removal patterns overlapped by 58–75% with clusters based on physicochemical properties, including Abraham descriptors and logKow. These results show that chemical descriptors can help estimate PPCP removal patterns and prioritize compounds for direct monitoring. Shotgun metagenomics was applied to air and potential source samples collected over four seasons at a dairy farm and a swine farm with contrasting ventilation designs. Clinically relevant ARGs, including putatively plasmid-associated ARGs, were detected in farm-associated air. Total ARG abundance, Rank I ARG abundance, and human health resistome risk were significantly higher at the mechanically ventilated swine farm than at the naturally ventilated dairy farm. Source-tracking analysis identified pen manure as the dominant inferred contributor to airborne ARGs at the dairy farm (64–81%) and dust as the largest inferred contributor among the sources sampled at the swine farm (53–67%). These findings suggest that manure and dust may be important sources of airborne ARGs at livestock farms. Because MGE annotation affects interpretation of ARG mobility, this work also addressed limitations of mobileOG-db, a widely used database of MGE-associated proteins developed at Virginia Tech. In the first version of the database, proteins inherited element-class labels from source databases, which could result in conflicting labels, and the database was composed primarily of sequences from cultured bacteria. mobileOG-db was reorganized into five non-overlapping element classes. A protein large language model classifier, mobileOG-ESM2, was then developed by fine-tuning ESM-2 on these classes. The classifier achieved Matthews correlation coefficients of 0.66 and 0.74 on test sets with maximum sequence identities of 40% and 80%, respectively. Applying the classifier to approximately 288 million proteins from public plasmid and viral databases expanded mobileOG-db 2.0 to 67.15 million protein entries. Together, these studies show how broad chemical analysis and shotgun metagenomics, combined with machine learning and improved reference databases, can provide more information from environmental samples for AMR surveillance. The computational approaches developed in this dissertation complement direct measurement and laboratory validation.