CLSSATP: Contrastive learning and self-supervised learning model for aquatic toxicity prediction.

Area of Science:

• Environmental Science
• Computational Chemistry
• Toxicology

Background:

• Rising aquatic compound concentrations threaten ecosystems and biodiversity.
• Assessing chemical impacts on aquatic organisms is crucial for environmental protection and sustainable development.
• Deep learning offers efficient, accurate, and generalizable alternatives to traditional toxicity testing.

Purpose of the Study:

• Introduce CLSSATP, a novel deep neural network for predicting organic toxicity.
• Leverage contrastive self-supervised learning for enhanced molecular representation and property prediction.
• Improve the accuracy and interpretability of aquatic toxicity assessments.

Main Methods:

• Developed CLSSATP, integrating self-supervised learning with molecular fingerprints and contrastive learning with molecular graphs.
• Employed dual-perspective learning to analyze molecular structure-property relationships.
• Conducted experiments and ablation studies to validate model performance.

Main Results:

• CLSSATP outperformed existing comparative methods in predicting organic toxicity.
• Ablation experiments demonstrated significant performance gains: 9.43% from the self-supervised module and 10.98% from the contrastive learning module.
• Model visualization confirmed the identification of key substructures influencing molecular properties, ensuring interpretability.

Conclusions:

• CLSSATP provides an effective and interpretable deep learning framework for aquatic toxicity assessment.
• The model's dual-perspective learning approach enhances understanding of molecular toxicity.
• This research offers a promising direction for future environmental risk assessment studies.