Multidimensional analysis of soil metal ecotoxicity and regional ecological risk prediction based on machine learning

Soil metal contamination threatens ecosystem stability and human health, yet risk management is hindered by the lack of thresholds and by models unable to capture multidimensional metal-soil-species interactions. Based on 1231 multidimensional data records compiled from the literature and complementary experimental sources, we developed a machine-learning framework to predict soil metal toxicity thresholds by integrating physicochemical properties of 12 metals (including copper, cadmium, zinc, etc.), key soil environmental factors, and species-specific responses, covering plants, invertebrates, and microorganisms (ten biological endpoints). Two encoding strategies, toxicity empirical value encoding and one-hot encoding were compared, and one-hot encoding combined with a random forest model achieved the best performance (R² = 0.82), with 98 % of predictions falling within 1.5-fold errors. Feature importance analysis shows soil properties as dominant contributors (42.7 %), particularly clay and cation exchange capacity (CEC), followed by metal attributes (32.8 %) and species sensitivity (24.5 %). Nonlinear interactions such as pH-clay were identified as key regulators of metal bioavailability. The trained model was subsequently applied to an independent dataset of 2139 soil samples from Europe and China, enabling prediction of EC₁₀ values for multiple metals, including emerging elements (tungsten (W), selenium (Se), tin (Sn), and thallium (Tl)), and reveals spatial risk patterns such as high-risk zones in Northern Europe and the North China Plain. These findings provide important support for large-scale, multidimensional soil metal risk assessment.