Artificial intelligence-powered new approach methodologies for assessing combined toxicity of chemical mixtures

Chemical contaminants in the environment invariably occur as complex mixtures, yet conventional risk assessment paradigms rely on single‑chemical evaluations that cannot capture emergent synergistic or antagonistic interactions. This review critically examines how artificial intelligence (AI) integrated with New Approach Methodologies (NAMs) is fundamentally reshaping mixture toxicity assessment. We directly compare four AI architectures traditional machine learning, Graph Neural Networks (GNNs), SMILES‑based Transformers, and the highly interpretable q‑RASAR framework, critically evaluating their respective abilities to model non‑additive mixture effects, handle data scarcity, and meet regulatory interpretability standards. We address critical blind spots in emerging contaminant mixtures, including microplastics acting as vectors for antibiotics and heavy metals, and nanomaterial‑driven acceleration of antibiotic resistance gene transfer. A central theme is the indispensable transition from statistical correlation to causal biological proof: we delineate the boundary between statistical explainability (SHAP values) and mechanistic explainability (Adverse Outcome Pathways), and we propose a closed‑loop validation strategy that combines in vitro organoid assays, AOP mapping, and AI‑optimized PBPK modeling. Finally, we provide a regulatory roadmap that emphasizes globally standardized validation protocols (e.g., the OECD (Q)SAR Assessment Framework), large‑scale chemistry‑biology interaction databases for zero‑shot mixture predictions, Bayesian uncertainty quantification, and legal frameworks for liability when using black‑box models.