Deep learning advances high-throughput toxicity screening of chemicals at multi-biological levels.

Deep learning (DL) has emerged as a powerful tool for modeling unstructured data, thereby improving prediction accuracy and expanding the application of machine learning (ML) in toxicity assessment. However, selecting suitable DL architectures and training methods for toxicity prediction remains challenging due to the lack of systematic comparisons regarding data types and modeling tasks across biological levels, which hinders the development of optimal models. To address these challenges, we review the current DL applications for predicting toxic events at four stages within the adverse outcome pathway framework: toxicophore-induced effects at the chemical exposure stage, activation of toxic pathways at the macro-molecular level (molecular initiating events), toxicogenomic responses at the cellular level (key events), and observable toxic effects (adverse outcomes) at the tissue/organ/individual levels. We compare the technical aspects of various DL methods for toxicity prediction and discuss how interpretability analyses can reveal the underlying molecular mechanisms and modes of toxic action. We also summarize current solutions to the challenges of increased data requirements and reduced interpretability of DL compared to traditional ML, and propose the development of a general environmental toxicological model. We hope that the interdisciplinary insights provided in this review can accelerate the development and application of new DL models in high-throughput toxicity screening, thereby advancing risk management strategies based on modes of toxic action.