Machine Learning to Predict Drug-Induced Liver Injury and Its Validation on Failed Drug Candidates in Development.

Area of Science:

• Pharmacology and Toxicology
• Computational Biology
• Drug Development

Background:

• Drug-induced liver injury (DILI) is a major hurdle in drug development, with current prediction methods showing limited human efficacy.
• Existing toxicological assessments often fail to accurately predict DILI risk for drug candidates.
• Novel approaches are needed to improve the prediction of DILI for pharmaceuticals under development.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting DILI risk using a large human dataset.
• To assess the performance of random forest (RF) and multilayer perceptron (MLP) models in DILI prediction.
• To validate the predictive capabilities of these models on drug candidates that failed clinical trials due to hepatotoxicity.

Main Methods:

• Leveraged a comprehensive human dataset to train machine learning models.
• Employed a 10-fold cross-validation strategy to evaluate model performance.
• Utilized an independent external test set, including drug candidates with known hepatotoxicity, for validation.

Main Results:

• Random forest (RF) achieved an average cross-validation accuracy of 0.631.
• Multilayer perceptron (MLP) demonstrated the highest Matthews Correlation Coefficient (MCC) of 0.245 during cross-validation.
• Both RF and MLP models successfully identified toxic drug candidates in the external validation set.

Conclusions:

• In silico machine learning models show significant potential for predicting DILI liabilities.
• These computational tools can aid in identifying hepatotoxic drug candidates early in the development pipeline.
• Machine learning offers a promising strategy to enhance the safety assessment of new drugs.