Machine-learning models based on biological ligand theory and quantitative ion character-activity relationship for predicting metal plant toxicity

Accurate prediction of metal phytotoxicity in soil-plant systems remains challenging, due to dynamic interactions between metal speciation, environmental variables and species-specific responses. To address this, this study incorporates mechanistic principles from the biotic ligand model (BLM) and quantitative ion character-activity relationship (QICAR) to construct a features-informed machine-learning framework (ML_BLM–QICAR) for predicting cross-species metal toxicity in heterogeneous soil conditions. Analyzing 2075 standardized experimental records spanning seven metals, five plants, and key environmental parameters (pH, competitive cations), twelve algorithms were evaluated. Of these, CatBoost demonstrated optimal performance (R² = 0.959, MSE = 0.176). Feature importance analysis of environmental factors, elemental characteristics and plant species revealed that metal additive amount (AMA, importance value: 0.745), pH and Mg²⁺/Ca²⁺ concentrations served as core driving factors. Notably, pH exhibited significant interaction networks with 19 other features. The proposed machine-learning (ML) model enables three-dimensional analysis of toxicity-influencing factors, outperforming traditional BLM and QICAR approaches. Application of this model further predicted toxicity thresholds (EC₁₀ = 7.493–1635.038 μM) for Co, Sb and the rare earth element Ce in four typical soil scenarios, with extreme values differing by as much as 218-fold. Additionally, the model validated the competitive adsorption mechanism of toxicity inhibition for Ce (R² = 0.924). This hybrid approach synergizes mechanistic theory with data-driven modeling, providing a transformative tool for rapid metal risk assessment and precision soil management in contaminated ecosystems worldwide.