Substructure-guided Deep Graph Learning in Molecular Toxicity Prediction.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Toxicology

Background:

• Machine learning and deep learning have advanced computational toxicity prediction.
• Challenges remain in handling data imbalance, missing labels, and model interpretability.
• Accurate toxicity prediction is crucial for efficient drug discovery.

Purpose of the Study:

• To develop a novel substructure-based deep graph learning architecture for toxicity prediction.
• To address data imbalance and missing labels in toxicity datasets.
• To enhance the interpretability of toxicity prediction models.

Main Methods:

• Developed a deep graph learning architecture incorporating functional groups into molecular graphs.
• Implemented strategies to manage missing labels and class imbalance.
• Utilized functional group-based feature importance analysis for interpretability.

Main Results:

• The developed model demonstrated strong performance in toxicity prediction tasks.
• The model effectively handled missing labels and class imbalance.
• Functional group analysis provided insights into toxicity mechanisms and improved model interpretability.

Conclusions:

• The substructure-based deep graph learning model offers a reliable tool for toxicity prediction.
• Enhanced model interpretability supports rational decision-making in drug development.
• This approach lays the groundwork for developing advanced toxicity prediction systems.