Discriminant models on mitochondrial toxicity improved by consensus modeling and resolving imbalance in training

Area of Science:

• Computational toxicology
• cheminformatics
• predictive modeling

Background:

• Chemical exposure poses risks, but experimental toxicity testing is challenging.
• Mitochondrial toxicity is linked to numerous diseases, necessitating efficient identification methods.

Purpose of the Study:

• To develop and validate quantitative structure-activity relationship (QSAR) models for predicting chemical mitochondrial toxicity.
• To improve the accuracy and efficiency of identifying chemicals that cause mitochondrial damage.

Main Methods:

• Employed five machine learning algorithms and twelve molecular fingerprints to build QSAR discriminant models.
• Utilized a threshold moving method to address imbalanced training data.
• Implemented a consensus strategy by averaging model probabilities to enhance prediction performance.

Main Results:

• Achieved correct classification rates of 81.8% in cross-validation and 88.3% in external validation.
• Identified specific substructures (phenol, carboxylic acid, nitro, arylchloride) as informative for predicting mitochondrial toxicity.
• Demonstrated that addressing data imbalance and using consensus models significantly improves prediction accuracy.

Conclusions:

• QSAR modeling offers a viable alternative to experimental testing for assessing chemical mitochondrial toxicity.
• Consensus modeling and threshold moving techniques are effective for improving predictive accuracy in toxicological assessments.
• The developed models provide a valuable tool for prioritizing chemicals for further investigation and mitigating health risks.