Integrating Functional Genomic Descriptors into Explainable Machine-Learning Models for Cross-Aquatic Species Prediction of Metal-Containing Nanomaterials Ecotoxicity

Understanding and predicting the ecotoxicity of nanomaterials (NMs) in untested aquatic species remains a major challenge due to the absence of quantitative and mechanistically meaningful species descriptors. Here, we propose a functional genomics-based species descriptor that captures molecular functions, antioxidant defenses, and detoxification capacities and evaluate its utility in improving deep-learning (DL) models for NM toxicity prediction. Using toxicity data for 14 metal-containing NMs across 31 aquatic organisms, we integrated the proposed descriptor into four supervised machine-learning algorithms, the models achieved a highly balanced and satisfactory predictive performance across different structures, with cross-validation Q2 values stabilizing around 0.61–0.65 for both random forest (RF) and DL architectures. Feature-importance analysis further showed that functional genomic traits contributed comparably to NM physicochemical properties and exposure conditions, indicating that the descriptor provides mechanistic relevance rather than functioning as a purely statistical feature set. This approach enables the quantitative characterization of species-specific sensitivity to NMs and substantially enhances cross-species generalizability of toxicity prediction. Consequently, the proposed genomics-based descriptor provides a scalable approach for hazard assessment of NMs in untested, rare, and data-deficient aquatic species of conservation concern.