Quantifying environment-dependent global antibiotic resistance gene risk patterns and mitigation scenarios using integrated machine learning and counterfactual scenario evaluation

Environmental antibiotic resistance genes (ARGs) pose major ecological and public health risks, yet environmentally conditioned risk patterns and mitigation potential remain poorly quantified. This study aimed to characterize global ARG risk patterns, identify associated environmental and agricultural factors, and evaluate potential mitigation scenarios using an integrated machine learning framework. Using 274,382 sample-ARG-environment records and 150 environmental and agricultural variables, quantile models characterized conditional risk index (RI) distributions and upper-tail risks, multi-model analysis assessed predictive variable importance, and DML estimated covariate-adjusted conditional associations. Heterogeneity was assessed across habitats, countries, and ARG categories, followed by directionally constrained counterfactual evaluation of multi-factor mitigation scenarios. Global ARG risk varied strongly with environmental conditions. In predictive importance analyses, winter temperature anomalies, manure nutrient loading, and pesticide application were important for upper-tail RI prediction. Separately, DML identified positive conditional associations of RI with insecticide application, cropland nutrient surplus, and manure loading, and negative associations with nutrient use efficiency. These associations varied across habitats, particularly soil, crop roots, freshwater, and biological waste environments. Under a predefined 50% predicted RI reduction target, nine-factor adjustment reduced mean standardized adjustment distance by 69.1% relative to single-factor adjustment, although outcomes varied among ARG categories. These model-based findings support global ARG risk characterization and management prioritization within the One Health framework, subject to empirical validation.