Predicting antibiotic and resistance gene dynamics in coastal urban waters under data scarcity via an integrated...

Antibiotics and antibiotic resistance genes (ARGs) in coastal urban waters pose increasing environmental and public health risks due to their persistence and their role in accelerating antimicrobial resistance (AMR), yet their dynamics remain difficult to predict under data-scarce conditions. This study developed an integrated multi-model framework that combines a process-based hydrological model, a modified LOADEST model, multiple machine learning approaches, and Shapley additive explanations (SHAP) for interpretability to predict antibiotics and AMR indicators (mainly ARGs) in three coastal zones of Singapore. Using readily available hydrometeorological datasets and limited routine water quality variables, the framework achieved robust performance, with the Long Short-Term Memory (LSTM) model performing best among the machine learning models. The LSTM-based integrated framework showed good predictive performance for low-censored antibiotics, as exemplified by clarithromycin (CLAR; NSE = 0.62 ± 0.13), and most AMR indicators across the three coastal zones (average NSE = 0.67 ± 0.10), whereas predictive performance decreased for highly censored antibiotics due to the limited information available for learning temporal variability. SHAP feature analysis showed that routine water quality variables and flow dominated model predictions, while meteorological factors were less influential, and SHAP temporal patterns further indicated that antibiotic responses to environmental drivers occurred on timescales comparable to catchment-to-coastal hydrological transport times, whereas ARGs consistently exhibited longer lagged responses across all regions. Incorporating predicted CLAR as an input feature in the macrolide-associated ARG model identified a positive association between CLAR and ARG only in the region with relatively high CLAR concentrations, and SHAP-derived interaction strengths showed that CLAR interacted most strongly with routine water quality variables, followed by hydrological factors, with weak interactions observed for meteorological variables. Overall, the proposed framework provides a practical and transferable hybrid modelling approach for predicting antibiotic and ARG dynamics in data-limited coastal environments.