Chemistry-Based Modeling on Phenotype-Based Drug-Induced Liver Injury Annotation: From Public to Proprietary Data.

Area of Science:

• Toxicology
• Computational Chemistry
• Machine Learning

Background:

• Drug-induced liver injury (DILI) is a significant safety issue leading to drug market withdrawal.
• Current DILI prediction models often rely on limited binary annotations from drug labels or case reports.
• Advancements in machine learning offer potential for improved in silico DILI prediction using chemical structures.

Purpose of the Study:

• To develop a more informative dataset for DILI prediction using a novel phenotype-based annotation of preclinical toxicology studies.
• To build and evaluate machine learning models for DILI prediction utilizing this enhanced dataset.
• To assess the impact of data source differences on model generalization.

Main Methods:

• Extracted hepatotoxicity information from in vivo preclinical toxicology studies using INHAND annotation.
• Created a dataset of 430 unique compounds with diverse liver pathology findings.
• Developed and trained DILI prediction models using compound fingerprints on the TG-GATEs dataset.
• Validated models using an external test set from Johnson & Johnson.

Main Results:

• Demonstrated successful prediction of DILI labels for TG-GATEs compounds.
• Quantified the impact of dataset differences on model generalization performance.
• Highlighted the utility of phenotype-based annotations for DILI prediction.

Conclusions:

• The novel phenotype-based annotation provides a more reliable dataset for machine learning-based DILI prediction.
• Machine learning models can effectively predict DILI using chemical structures and preclinical data.
• Understanding dataset differences is crucial for robust model generalization in DILI prediction.