Multigenerational Toxicity Prediction Using Graph Neural Networks and Biotransformation Analysis

Predicting how chemical toxicity evolves during metabolism remains a central challenge in environmental toxicology. This study presents an integrated computational framework combining descriptor-based QSAR modeling, graph neural networks, and biotransformation pathway analysis to capture the dynamic nature of the chemical risk. An XGBoost model using PaDEL descriptors achieved high accuracy (R2 = 0.969), while a newly developed Graph Sequence Attention Network (GSAT) yielded comparable results (R2 = 0.907, mean absolute error (MAE) ≈ 0.30) with a 98.9% reduction in computation time, enabling high-throughput screening and substructure-level interpretation. Integrating GSAT with EnviPath-python and KEGG validation enabled the simulation of multigenerational toxicity for 12 priority pollutants. Across transformation networks, 68.2% of the transformation steps increased the toxicity, resulting in an average of a 15.4-fold increase. The most detoxifying pathways included polynuclear aromatic dioxygenation, decarboxylation to CO2, and the conversion of halogenated muconate to succinate, whereas alcohol oxidations consistently enhanced toxicity. Notably, 91.6% of the halogenated pathways showed net toxification, with 78.3% reaching peak levels during steps 3–4, indicating temporary periods of maximum hazard before detoxification. These findings reveal that metabolic transformations often increase toxicity and emphasize that pathway level rather than static toxicity assessment is essential for accurate environmental risk evaluation and the rational design of bioremediation strategies.