Predictors Recover Most of the Metagenomic Signal for Antibiotic Resistance Gene Occurrence: A Cross-City Test of Geographic Transferability in Urban Wastewater

Abstract

Antibiotic resistance genes (ARGs) travel from cities into rivers and coastal waters through wastewater treatment plants. Monitoring them normally requires metagenomic sequencing, which is sufficiently costly that most utilities can sample only occasionally. Weather and location data are freely available on a daily basis for virtually any location worldwide, making them attractive predictors for identifying where limited sequencing resources should be prioritized. However, whether such models generalize to previously unseen cities remains largely untested, as most published studies train and evaluate models within the same catchments, thereby assessing interpolation rather than geographic transferability. To address this gap, we paired 235 wastewater metagenomes collected from five European cities with 23 abiotic predictors spanning geospatial, meteorological, hydrological, radiative, and temporal domains. Model performance was evaluated using leave-one-group-out (LOGO) cross-validation, in which all samples from one city were withheld for testing while the remaining cities were used for training. CatBoost achieved a median LOGO ROC-AUC of 0.929 and a median F1-score of 0.750. Using freely available environmental reanalysis predictors (meteorological, hydrological, radiative, geospatial, and temporal) - without any metagenomic sequencing - CatBoost achieved a mean ROC-AUC of 0.722, recovering 78% of the predictive performance of the full omics-integrated model. Removing latitude and longitude reduced ROC-AUC by only 0.003, whereas replacing random cross-validation with city-wise validation reduced ROC-AUC by 0.052. Predictive performance varied across ARG classes, ranging from a ROC-AUC of 0.981 for {beta}-lactam resistance genes to 0.762 for glycopeptide resistance genes. These findings demonstrate that freely available environmental reanalysis predictors - spanning meteorological, hydrological, radiative, geospatial, and temporal domains - recover 78% of the predictive signal for antibiotic resistance gene occurrence in urban wastewater, allowing scarce sequencing capacity to be directed to the catchments where it changes a decision.