Multi-task aquatic toxicity prediction model based on multi-level features fusion.

Area of Science:

• Environmental Science
• Computational Chemistry
• Toxicology

Background:

• Organic compounds pose a significant threat to aquatic organisms.
• Assessing aquatic toxicity is crucial for environmental protection and understanding ecological impacts.
• Deep learning offers superior accuracy and speed for toxicity prediction compared to traditional methods.

Purpose of the Study:

• To introduce ATFPGT-multi, an advanced multi-task deep neural network for predicting organic compound toxicity.
• To evaluate the efficacy of multi-task learning in aquatic toxicity prediction.

Main Methods:

• ATFPGT-multi integrates molecular fingerprints and graphs to characterize organic molecules.
• The model simultaneously predicts acute toxicity across four fish species.
• Cross-validation was employed to assess performance and generalization ability.

Main Results:

• ATFPGT-multi demonstrated superior performance over single-task models (ATFPGT-single) across four fish datasets.
• The model achieved higher accuracy and reliability compared to previous algorithms.
• Attention scores from ATFPGT-multi identified key molecular fragments linked to fish toxicity, showcasing model interpretability.

Conclusions:

• ATFPGT-multi provides a robust framework for advancing aquatic toxicity assessment.
• The model's open-source availability facilitates further research and application in environmental protection.