Developing a GNN-based AI model to predict mitochondrial toxicity using the bagging method.

Area of Science:

• Biochemistry
• Toxicology
• Computational Chemistry

Background:

• Mitochondrial toxicity is a critical factor in drug development, linked to various toxicities like hepatotoxicity.
• Current predictive models offer binary outcomes without identifying contributing structural elements.
• Early-stage screening for mitochondrial toxicity is essential to prevent late-stage failures.

Purpose of the Study:

• To develop an artificial intelligence (AI) model for predicting mitochondrial toxicity from chemical structures.
• To enable visualization of structural alerts contributing to predicted toxicity.
• To enhance early drug discovery by providing detailed toxicity insights.

Main Methods:

• Utilized the kMoL software library with a graph neural network approach for chemical structure learning.
• Employed the integrated gradient method for visualizing substructures associated with positive toxicity predictions.
• Addressed dataset imbalance using the bagging method to improve model performance.

Main Results:

• Developed an AI model with high predictive performance, achieving an F1 score of 0.839.
• Successfully visualized specific substructures contributing to mitochondrial toxicity predictions.
• Demonstrated the model's capability in identifying potential mitochondrial toxicants.

Conclusions:

• The developed AI model accurately predicts mitochondrial toxicity based on chemical structures.
• The model's visualization feature aids in understanding the structural basis of toxicity.
• This tool can significantly contribute to the early-stage screening of drug candidates for mitochondrial safety.