Machine learning assessment of fungicide ecotoxicity using multi-species biomarkers and an early risk warning system

The widespread use of fungicides has raised significant concerns regarding their ecotoxicological risks. However, most existing studies are limited to single compounds, species, or endpoints. This study developed interpretable machine learning models to assess the toxicity of fungicides across soil and aquatic organisms, incorporating 30 mechanistic biomarkers (9 for soil organisms and 21 for aquatic organisms). Using 21 algorithms, 672 classification models were constructed. Gradient Boosting and Tree-based methods outperformed other approaches, achieving Receiver Operating Characteristic - Area Under the Curve values of 0.990-1.000 (training) and 0.861-1.000 (testing). Internal validation (leave-one-out cross-validation, repeated 5-fold validation, and bootstrap resampling) and external validation (independent test set) collectively confirmed the model's robust predictive capability and strong generalization performance. SHapley Additive exPlanations analysis identified malondialdehyde (MDA) and reactive oxygen species (ROS) as the most influential biomarkers. An early-risk warning tool with a graphical user interface was developed using MDA and ROS for rapid risk assessment. This study establishes a mechanism-driven framework that leverages machine learning and key biomarkers to not only advance the predictive assessment of fungicide ecological risks but also provide a scientific basis for proactive early risk warning and targeted management.