TRIDENT-2 predicts chemical toxicity across Eukaryota from 560,780 assays with average median absolute error of 1.76 to 3.80

The authors built and trained TRIDENT-2, a multimodal artificial intelligence model for predicting chemical toxicity across Eukaryota. Compared with toxicity prediction aimed at narrower sets of species or chemicals, the model explicitly targets evolutionarily diverse eukaryotic species and jointly incorporates chemical, biological, and experimental information. The abstract reports training on 560,780 toxicity assays covering 82,775 chemicals, 6,793 species, and multiple exposure scenarios; the loaded text is the abstract only and does not include model architecture, validation splits, or baseline comparisons.

TRIDENT-2 reports an average median absolute error of 1.76 to 3.80 for toxicity prediction across Eukaryota. This error range provides a citable quantitative accuracy reference for cross-species toxicity prediction, which has been difficult to obtain for data-scarce species. The error values come directly from the abstract's stated results; because the loaded text does not give the error unit, the composition of the evaluation set, or statistical uncertainty, readers need the original article to confirm these details.

By jointly learning from chemical, biological, and experimental information, the model remains accurate across broad chemical and taxonomic distances, enabling toxicity assessment for species and chemicals beyond current experimental evidence. This extends the reach of toxicity assessment from species and chemicals already covered by experiments to species-chemical combinations where evidence is still missing. The abstract states that the model "remains accurate across broad chemical and taxonomic distances," but does not specify how distance is measured or the exact setup of the extrapolation tests.

The authors conclude that artificial intelligence can help overcome longstanding data limitations in ecotoxicology, supporting improved decision-making and reducing chemical impacts on biodiversity and ecosystems. This links a methodological advance to biodiversity protection and management needs rather than only improving accuracy on a single prediction task. This is a concluding statement at the abstract level and reflects the authors' judgment about application prospects; the loaded text contains no empirical evaluation of decision-making use or ecological impact.

The work addresses ecotoxicology and chemical risk assessment settings, for users who need toxicity estimates for species or chemicals lacking experimental data, such as research and management related to ecological risk assessment and biodiversity protection. Its value proposition is to make toxicity assessment available for species and chemicals beyond current experimental evidence, thereby supporting decision-making. Whether it applies to specific determinations under a particular regulatory framework depends on validation details not presented in the abstract.

The loaded text is the abstract only and contains no figures, model structure, data splits, baseline comparisons, or error units, so the precise meaning and comparability of 1.76 to 3.80 cannot be judged, nor can the measurement behind "remains accurate across broad chemical and taxonomic distances" be confirmed. Readers would still watch how predictions perform on taxa with sparse training data, whether error is consistent across exposure scenarios, and how model outputs would connect to existing ecological risk assessment workflows. These are open questions for the original article to address.