Deep Graph Learning with Property Augmentation for Predicting Drug-Induced Liver Injury.

Area of Science:

• Computational chemistry
• Pharmacology
• Toxicology

Background:

• Drug-induced liver injury (DILI) is a critical safety concern in drug development.
• Accurate DILI prediction is challenging due to complex testing and limited annotated data.
• Early-stage in silico screening can reduce drug development costs by identifying high-risk candidates.

Purpose of the Study:

• To develop an accurate computational method for predicting DILI properties.
• To address the challenge of limited annotated data for DILI prediction models.
• To improve the efficiency of early-stage drug discovery by filtering potential DILI-causing drug candidates.

Main Methods:

• Application of traditional machine learning and graph-based deep learning techniques.
• Development of a property augmentation strategy to overcome data scarcity.
• Extensive experimental validation using various cross-validation strategies (random, leave-one-out, scaffold splitting).

Main Results:

• The proposed method significantly outperforms existing baseline models for DILI prediction.
• Achieved high accuracy rates: 81.4% (random splitting), 78.7% (leave-one-out), and 76.5% (scaffold splitting).
• Property augmentation effectively mitigated the data scarcity issue for deep learning models.

Conclusions:

• The developed computational approach offers a robust solution for predicting DILI.
• This method can aid in early-stage drug discovery by effectively filtering drug candidates with high DILI risk.
• The property augmentation strategy is a valuable technique for improving predictive model performance with limited datasets.