Multi-Task Graph Convolutional Network Model with Improved Performance and Broad Applicability Domains for Identifying Aquatic Toxic Chemicals

High-throughput computational models are essential for regulatory screening of large chemical inventories for aquatic toxicity, while their performance is frequently constrained by limited datasets, inadequate exploitation of inter-endpoint correlations, and lack of applicability domain (AD) characterization. This study curated a regulation-oriented aquatic toxicity dataset comprising 17,514 chemicals across six ecotoxicological endpoints. An end-to-end multi-task learning graph convolutional network (MTL-GCN) was constructed to enable joint modeling of multiple toxicity endpoints. Relative to single-task learning (STL-GCN), the MTL-GCN improved the area under the receiver operating characteristic curve by 0.4~13.4% across the six endpoints. A weight similarity parameter (PWS) was calculated, indicating that MTL effectively captures intrinsic correlations among endpoints. The integrated attention mechanism enhanced interpretability by identifying aquatic toxicophores. AD characterization method based on the structure-activity landscape (ADSAL) analysis confirmed that MTL-GCN improves in-domain screening accuracy while maintaining comparable coverage to STL-GCN. When applied to 783,681 inventory chemicals, the combined MTL-GCN and ADSAL framework achieved reliable screening with 94.7% of chemicals located in the AD of at least one endpoint. The constructed dataset and the MTL-GCN framework establish benchmarks for aquatic toxicity prediction and provide a robust method for large-scale ecological hazard assessment.