Machine learning-based prediction of antibiotic resistance gene distribution in agricultural soils under different climate change scenarios.
Source: PubMed, NCBI / U.S. National Library of Medicine
Antibiotic resistance genes (ARGs) in agricultural soils represent a major public health concern, as climate change is believed to augment their dissemination and abundance. Understanding the impact of future climate change scenarios on ARG abundance is essential to implement predictive and proactive One Health strategies. In this study, a total of 2301 soil samples from 67 countries across six continents were compiled from three global metagenome databases, namely NCBI SRA, MG-RAST, and JGI IMG/M. Six machine learning models, namely LightGBM, XGBoost, Random Forest, Support Vector Machines, Deep Neural Networks, and Logistic Regression, were used to predict ARG distribution patterns in agricultural soils, and their performance was evaluated using stratified 10-fold cross-validation with metrics such as AUC-ROC, precision, recall, F1 score, and Matthews Correlation Coefficient. WorldClim 2.1 and CMIP6 models were used to project ARG distribution under three Representative Concentration Pathway scenarios, namely RCP 2.6, RCP 4.5, and RCP 8.5, for the years 2050 and 2070. The LightGBM model achieved the best predictive performance, with an AUC-ROC of 0.957 (95% CI: 0.951-0.963), substantially higher than that of the other models, while the Deep Neural Networks model achieved an AUC-ROC of 0.891. The LightGBM model demonstrated high stability across cross-validation folds, with minimal fold-to-fold variance, defined as the standard deviation of AUC-ROC scores across the 10 folds
Abstract
Antibiotic resistance genes (ARGs) in agricultural soils represent a major public health concern, as climate change is believed to augment their dissemination and abundance. Understanding the impact of future climate change scenarios on ARG abundance is essential to implement predictive and proactive One Health strategies. In this study, a total of 2301 soil samples from 67 countries across six continents were compiled from three global metagenome databases, namely NCBI SRA, MG-RAST, and JGI IMG/M. Six machine learning models, namely LightGBM, XGBoost, Random Forest, Support Vector Machines, Deep Neural Networks, and Logistic Regression, were used to predict ARG distribution patterns in agricultural soils, and their performance was evaluated using stratified 10-fold cross-validation with metrics such as AUC-ROC, precision, recall, F1 score, and Matthews Correlation Coefficient. WorldClim 2.1 and CMIP6 models were used to project ARG distribution under three Representative Concentration Pathway scenarios, namely RCP 2.6, RCP 4.5, and RCP 8.5, for the years 2050 and 2070. The LightGBM model achieved the best predictive performance, with an AUC-ROC of 0.957 (95% CI: 0.951-0.963), substantially higher than that of the other models, while the Deep Neural Networks model achieved an AUC-ROC of 0.891. The LightGBM model demonstrated high stability across cross-validation folds, with minimal fold-to-fold variance, defined as the standard deviation of AUC-ROC scores across the 10 folds (SD = 0.008). SHAP feature importance analysis identified soil temperature, pH, and organic carbon content as the top three factors influencing ARG relative abundance, with SHAP values of 0.342, 0.287, and 0.251, respectively. Annual precipitation and soil moisture level were also identified as significant contributors to ARG distribution. SHAP dependency plots revealed critical thresholds for ARG relative abundance, with a sharp increase observed independently when soil temperature exceeds 18 °C and when soil pH drops below 6.5. Furthermore, a non-linear accelerating increase in ARG abundance risk was observed as climate change intensity worsened across scenarios. Projections for future climate change scenarios indicate a potential 34.7% increase in high-risk ARG zones by the year 2070, with the largest changes expected in South Asia, Sub-Saharan Africa, and Mediterranean regions. Paired t-tests revealed significant differences in performance among all models (p < 0.001). These findings demonstrate that gradient-boosting methods such as LightGBM outperform deep learning approaches for ARG prediction from soil microbiome data, offering higher accuracy and interpretability. As climate change is projected to increase ARG risks in a non-linear manner, the development of climate-adaptive agricultural practices and global surveillance systems is urgent. This framework provides actionable risk-mapping tools to support precision farming and region-specific policy interventions within the One Health approach.