Multimodal Integration of Chemical and Biological Descriptors for Cross-Species Prediction of Fish Acute Toxicity

The expanding diversity of synthetic chemicals is increasing ecological risk, yet many predictive models rely on single-species data and inadequately capture interspecies variability in sensitivity. We propose a generalized toxicity prediction framework (GTGT) that predicts species-resolved acute toxicity (log₁₀LC₅₀; LC₅₀ in mg/L) from chemical descriptors, exposure duration, and species features─taxonomy embeddings and mitochondrial Cytochrome <i>b</i> (<i>cytb</i>) sequence embeddings─within a unified deep-learning architecture. Using a dataset of 2860 compounds and 297 fish species, GTGT outperformed representative state-of-the-art models, achieving external-test <i>R</i>² = 0.83 and RMSE = 0.49. Ablation analyses show that chemical and exposure descriptors provide baseline performance, whereas biological features are critical to capture interspecies susceptibility. Comparative analyses further indicate that taxonomy embeddings encode hierarchical evolutionary relationships, while <i>cytb</i> sequences capture molecular divergence, providing complementary information for robust cross-species prediction. We also provide a web platform for single- and multicompound predictions across multiple fish species, enabling model-based species sensitivity distribution (SSD) curves. This framework links chemical, biological, and exposure dimensions to support SSD parametrization and derivation of protective thresholds for comparison with environmentally relevant exposures, rather than serving as a direct risk metric.