Multi-Endpoint Acute Toxicity Assessment of Organic Compounds Using Large-Scale Machine Learning Modeling

Area of Science:

• Computational toxicology
• Machine learning in drug discovery
• Predictive modeling for chemical safety

Background:

• Toxicity testing increasingly relies on alternative methods like computational and machine learning (ML) approaches.
• Developing accurate predictive models is hindered by complex and scarce biomedical data.
• Multicondition descriptors (MCDs) combined with nonlinear ML offer a robust solution for integrating diverse assay data.

Purpose of the Study:

• To develop a Quantitative Structure-Toxicity Relationship (QSTR) model using multicondition descriptors (MCDs).
• To evaluate the predictive performance of single-task, multi-endpoint ML models and Convolutional Neural Networks (CNNs).
• To identify key structural features influencing compound toxicity.

Main Methods:

• Application of multicondition descriptors (MCDs) to a large dataset (>80,000 compounds, 59 endpoints).
• Development and comparison of single-task multi-endpoint ML models.
• Utilizing Convolutional Neural Networks (CNNs) for a novel data analysis approach.

Main Results:

• The use of MCDs significantly improved model performance.
• CNN-1D models combined with MCDs achieved the best predictive accuracy (R²train = 0.93, R²ext = 0.70).
• Key toxicity-associated structural features identified include VSA, nN+, S-P fragments, ionization potential, and C-N fragments.

Conclusions:

• MCDs enhance the robustness and accuracy of QSTR models.
• CNN-1D models integrated with MCDs provide a powerful tool for toxicity prediction.
• The developed models facilitate rapid toxicity assessment of novel chemical compounds.