Artificial intelligence in toxicology: current advances, challenges and future directions

Toxicology has traditionally relied on in vivo animal studies and in vitro assays to assess the safety of drugs, chemicals and environmental contaminants. These approaches are constrained by cost, duration, ethical concerns and difficulties in cross-species extrapolation. Artificial intelligence (AI) offers an alternative paradigm in which safety-relevant signals are extracted from large, heterogeneous datasets. This review traces the development of AI in toxicology from rule-based expert systems to contemporary deep learning, large language model and multimodal architectures. It critically evaluates the evidence base for AI-based toxicity prediction, assesses current challenges and considers the trajectory toward next-generation risk assessment. A structured literature search was conducted in PubMed/MEDLINE, Web of Science and Scopus (2000-2026), supplemented by guidance documents from OECD, FDA, EMA, EFSA and EPA. Studies were selected based on methodological quality, availability of external validation data and regulatory relevance. Graph neural networks, multi-task deep learning and related AI approaches have shown competitive performance in selected benchmark studies of drug-induced liver injury (DILI), hERG cardiotoxicity and Ames mutagenicity, although performance varies substantially across datasets and validation settings. Explainable AI frameworks such as SHAP are aligning model outputs with adverse outcome pathways (AOPs). Federated learning enables privacy-preserving multi-institutional collaboration, as demonstrated by the MELLODDY consortium. Critical gaps remain: external validation is inconsistently reported, endpoint-specific models struggle to generalise and hallucination in large language models poses unresolved regulatory risks. AI is becoming an increasingly important component of toxicological science, although its readiness for routine application varies substantially across methods and endpoints. Realising this potential requires adherence to the TREAT validation principles and sustained regulatory engagement. AI models augment, but cannot yet replace, experimental toxicology.