To assess the environmental risk of chemicals, extensive freshwater ecotoxicity data and prediction models have been established. However, due to the scarcity of saltwater ecotoxicity data, no model was reported to predict toxicity endpoints across various marine species. To answer this challenge, we integrated multitask deep learning with cross-domain knowledge-sharing mechanism to develop MarineTox predictor, a novel computational tool for facilitating end-to-end prediction of 31 saltwater toxicity tasks for 26 organisms spanning five phyla. By sharing substructure-based features from freshwater ecotoxicity data, MarineTox predictor effectively improved prediction capacity for 18 low-resourced tasks (R2 in validation set of 0.5-0.93), achieving an R2 improvement of up to 140% compared to a model trained solely on saltwater ecotoxicity data. To decode structure-dependent toxicity mechanisms, we constructed an interaction network between multiple species and substructures for identifying six key substructures with toxicity to specific marine organisms. Using the predicted ecotoxicity data, we systematically derived environmental hazard thresholds for ∼68,000 chemicals to decipher a holistic view of aquatic hazards, determining 902 chemicals of toxicity concern to the marine ecosystem. Ultimately, an online library platform (https://marinetox-predictor-dlut.streamlit.app/) covering 1.2 million records of ecotoxicity data and hazard thresholds was established, supporting marine ecological risk assessment of diverse chemicals.
Zhu et al. (2026) studied this question.