Quantitative Toxicity Prediction Using Topology Based Multitask Deep Neural Networks.

Area of Science:

• Computational chemistry
• Toxicology
• Algebraic topology

Background:

• Understanding chemical toxicity is vital for human health and environmental safety.
• Quantitative toxicity analysis is the emerging standard in the field.
• Existing methods may not fully capture complex molecular information.

Purpose of the Study:

• To introduce Element-Specific Persistent Homology (ESPH) for quantitative toxicity prediction.
• To explore the representability and predictive power of ESPH for small molecules.
• To develop and validate advanced machine learning models for toxicity analysis.

Main Methods:

• Utilized algebraic topology to develop Element-Specific Persistent Homology (ESPH) for molecular representation.
• Developed ancillary descriptors based on physical models.
• Integrated topological and physical descriptors with machine learning algorithms (DNN, RF, GBDT).
• Proposed a topology-based multitask strategy for handling diverse data set sizes.

Main Results:

• ESPH effectively retains crucial chemical information, providing a unique molecular representation.
• The combination of topological and physical descriptors with machine learning demonstrated strong predictive performance.
• The proposed topological learning methods outperformed state-of-the-art approaches in quantitative toxicity analysis.
• Validated on four benchmark quantitative toxicity data sets.

Conclusions:

• ESPH is a powerful tool for quantitative toxicity prediction, offering superior molecular representation.
• The integration of topological descriptors with machine learning provides a robust framework for toxicity assessment.
• The developed methods represent a significant advancement in computational toxicology.