Machine Learning-Based Prediction of Drug-Induced Hepatotoxicity: An OvA-QSTR Approach

Area of Science:

• Computational toxicology
• Pharmacovigilance
• Drug safety assessment

Background:

• Drug-induced hepatotoxicity, or drug-induced liver injury (DILI), is a significant cause of patient morbidity and mortality.
• The LiverTox database classifies drugs into risk categories (A-E) to mitigate liver toxicity.
• Existing methods for assessing drug-induced liver damage often involve animal testing.

Purpose of the Study:

• To develop a predictive model for assessing the liver damage potential of new drug molecules.
• To classify drugs into LiverTox risk categories without relying on experimental animals.
• To provide a computational tool for premarketing risk evaluation of pharmaceutical-induced liver damage.

Main Methods:

• Utilized a dataset of 678 organic drug molecules from the LiverTox database.
• Implemented a one-vs-all quantitative structure-toxicity relationship (OvA-QSTR) model.
• Employed Bayesian Network (BayesNet) for model implementation and analysis.

Main Results:

• The OvA-QSTR models demonstrated strong performance with precision-recall curve (PRC) areas ranging from 0.718 to 0.869.
• The models successfully predicted the LiverTox risk category for drugs with unknown liver toxicity potential.
• The developed models offer reliable predictions for pharmaceutical-induced liver damage.

Conclusions:

• The OvA-QSTR models provide a robust, non-animal-based method for evaluating drug-induced liver damage potential.
• This approach facilitates reliable premarketing risk assessment for drug safety.
• The study offers predictions for various risk levels associated with drug-induced liver injury (DILI).