OpenTox Virtual Conference 2026

Machine Learning–Driven QSTR Modeling of Heterocyclic Fungicide Ecotoxicity Across Aquatic Trophic Levels

Machine-learning approaches are increasingly used in predictive ecotoxicology, yet their performance may vary substantially among biological endpoints. This work examined the ability of three nonlinear QSTR approaches -support vector regression (SVR), Gaussian process regression (GPR), and multilayer perceptron neural networks (MLP-NN)- to predict the acute aquatic toxicity of structurally diverse heterocyclic fungicides. Separate models were evaluated for Daphnia magna, Oncorhynchus mykiss, and Pseudokirchneriella subcapitata, representing distinct aquatic trophic levels. Model performance was compared using multiple statistical criteria integrated through a multi-criteria decision-making framework. The analysis revealed clear differences in algorithm performance among endpoints. GPR provided the strongest overall prediction for D. magna, whereas MLP-NN showed the most balanced performance for the fish and algal endpoints. For D. magna, the GPR model achieved a test R² of 0.665. These findings indicate that model selection in aquatic QSTR should consider endpoint-specific response patterns rather than assuming that a single machine-learning algorithm will perform optimally across different species. These findings highlight the potential of comparative machine-learning approaches for endpoint-specific aquatic toxicity prediction and their possible contribution to future ecotoxicological assessments of fungicides.