Cross-modal attention deep learning reveals how transformation products inherit life-cycle risks from parent antibiotics: Insights for environmental, ecological, health, and AMR risks.

The environmental, ecological, antimicrobial resistance (AMR), and human health risks of antibiotic transformation products (TPs) in aquatic environments-especially high-similarity transformation products (HSTPs) containing active groups of parent antibiotics-remain poorly understood. This study assesses the ecological (EcoR), environmental (EnvR), human health (HuhR), and AMR risks of 141 parent antibiotics, introducing a few-shot cross-modal attention deep learning (CMA-DL, based on molecular graphs (GINConv or GATConv) and fused molecular fingerprints) model to identify multi-category antibiotic life-cycle risk (LCR) priorities. Analysis revealed that fluoroquinolones (82.14%) and sulfonamides (90.91%) dominate the high and low LCR priority groups, with particularly notable HuhR and AMR risk for both. Compared to traditional machine learning models (random forest, extreme gradient boosting, and support vector machines), the CMA-DL model demonstrated superior predictive accuracy and robustness. Interpretability analysis revealed that HSTPs containing more nitrogen, carboxyl, carbonyl, hydroxyl, and halogen groups were more likely to be classified as high LCR priority. The model predicted that around 90% of HSTPs retained similar LCR risk levels (high or medium) as their parent antibiotics, with approximately 2% of HSTPs showing higher LCR risks, primarily due to transformations like N-dealkylation, N-hydroxylation, N-acetylation, or amine chlorination. This study aims to provide novel insights into understanding the complex health and toxicity risks of antibiotics and their HSTPs, while offering theoretical support for developing effective management strategies for antibiotics and their HSTPs in aquatic environments (including priority antibiotic classes, LCR types, and antibiotic transformation pathways).