Comprehensive hepatotoxicity prediction: ensemble model integrating machine learning and deep learning.

Area of Science:

• Computational toxicology
• Drug discovery and development
• Machine learning in pharmacology

Background:

• Chemicals can cause acute liver injuries, a significant health concern.
• Assessing compound safety is complex and costly.
• In silico methods can identify drug candidate risks early, reducing development expenses.

Purpose of the Study:

• To develop accurate Quantitative Structure-Activity Relationship (QSAR) models for predicting chemical hepatotoxicity.
• To integrate machine learning (ML) and deep learning (DL) algorithms for enhanced prediction accuracy.
• To mitigate drug development costs by enabling early identification of liver injury risks.

Main Methods:

• Developed QSAR models for hepatotoxicity prediction using an ensemble strategy.
• Integrated diverse ML and DL algorithms with various molecular descriptors and fingerprints.
• Utilized feature selection and hybrid ensemble approaches for model optimization.
• Trained models on a dataset of 2588 chemicals and drugs, divided into training (80%) and test (20%) sets.

Main Results:

• The voting ensemble classifier achieved optimal performance with 80.26% prediction accuracy, 82.84% AUC, and over 93% recall.
• Ensemble methods like bagging and stacking also showed strong performance.
• The model demonstrated superior reliability compared to existing published models through external validation and cross-validation.

Conclusions:

• The proposed ensemble model provides a reliable and high-performing method for predicting chemical-induced liver damage.
• This approach offers a dependable assessment of hepatotoxicity risk for chemicals and drugs.
• Facilitates early-stage risk assessment in drug development.