Assessing chemical toxicity across Eukaryota using multimodal transformers

Biodiversity worldwide faces threats from chemical pollution, yet comprehensive toxicity data are lacking for vast numbers of species and chemicals, severely hindering our understanding of ecological impacts. In this study, researchers introduce TRIDENT-2, an advanced artificial intelligence model designed to predict chemical toxicity across a wide range of eukaryotic species.

This model is trained on an extensive dataset of 560,780 toxicity assays, encompassing 82,775 chemicals, 6,793 species, and various exposure scenarios.

TRIDENT-2 demonstrates remarkable accuracy in predicting toxicity across the diverse group of Eukaryota, with an average median absolute error ranging from 1.76 to 3.80. The model achieves this accuracy by integrating information from chemical properties, biological characteristics, and experimental data. This multimodal approach allows TRIDENT-2 to maintain high predictive capability even when dealing with significant differences in chemical composition and taxonomic classifications, thus enabling the assessment of toxicity for species and chemicals that lack current experimental evidence.

The findings of this research highlight the potential of artificial intelligence to address longstanding data limitations in the field of ecotoxicology. By overcoming these challenges, TRIDENT-2 offers a promising tool for enhancing decision-making processes and mitigating chemical impacts on biodiversity and ecosystems. This advancement underscores the transformative role of AI in addressing critical environmental issues.