MCPNET: Development of an interpretable deep learning model based on multiple conformations of the compound for...

Area of Science:

• Computational toxicology
• cheminformatics
• deep learning

Background:

• Predicting toxicological endpoints is crucial for drug development.
• Deep learning models offer promise but face challenges in accuracy and interpretability.
• Identifying a compound's bioactive conformation is difficult without experimental data.

Purpose of the Study:

• To develop an interpretable deep learning framework for predicting developmental toxicity.
• To address the challenge of determining bioactive conformations in deep learning models.
• To improve the accuracy of toxicological endpoint prediction using 3D molecular representations.

Main Methods:

• Developed the Multi-Conformation Point Network (MCPNET) deep learning framework.
• Utilized electrostatic potential distributions on vdW surfaces of multiple compound conformations.
• Applied 3D multi-conformational surface point clouds for feature extraction.

Main Results:

• Achieved 85% accuracy on classification and R² of 0.66 on regression tasks for developmental toxicity prediction.
• Outperformed traditional machine learning and other deep learning models.
• Demonstrated model interpretability through component visualization to understand compound mechanisms.

Conclusions:

• MCPNET accurately predicts developmental toxicity using 3D molecular representations.
• The model's interpretability aids in understanding compound action mechanisms.
• MCPNET shows potential for advancing toxicological predictions in drug discovery.