Machine learning-based prediction of phenanthrene accumulation and toxicity in earthworms across soils

Phenanthrene (PHE), a representative polycyclic aromatic hydrocarbon, readily bioaccumulates in soil organisms and poses substantial ecological risks. However, accurately predicting PHE toxicity across heterogeneous soils remains challenging due to complex soil-organism interactions. In this study, a machine learning (ML) framework was developed to predict PHE toxicity. Laboratory toxicity assays in 24 typical Chinese soils revealed a strong linear relationship between internal concentration (C_e) of PHE and earthworm fecundity, described by linear regression (LR). A comprehensive database encompassing soil physicochemical properties and PHE concentration (C_s) was constructed to support C_e prediction. Among 11 ML algorithms evaluated, CatBoost exhibited superior performance (R² = 0.8), identifying C_s as the primary determinant of C_e, followed by key soil properties (pH, soil organic matter). The integrated CatBoost-LR model accurately predicted toxicity (R² = 0.81) within external verification and enabled spatial assessment of PHE-associated ecological risks across China. These findings demonstrated that ML can quantitatively link soil properties to pollutant bioaccumulation and toxicity, providing a powerful, data-driven tool for ecological risk assessment of organic contaminants in complex terrestrial environments.