Comparing Machine Learning Algorithms for Predicting Drug-Induced Liver Injury (DILI).

Area of Science:

• Pharmacology
• Toxicology
• Computational Chemistry

Background:

• Drug-induced liver injury (DILI) is a major cause of drug withdrawal.
• The FDA's DILIRank database classifies DILI severity and potential.
• Existing DILI prediction models utilize FDA data and in vitro assays.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting DILI.
• To integrate DILI data with physicochemical and biopharmaceutics properties.
• To create a predictive tool for early identification of hepatotoxic compounds.

Main Methods:

• Utilized DILIRank, in vitro DILI data, and Biopharmaceutics Drug Disposition Classification System data.
• Developed Bayesian machine learning models using Assay Central software.
• Assessed model performance via 5-fold cross-validation and an external test set.

Main Results:

• The best Bayesian model, using DILIRank data, achieved an ROC of 0.814.
• Achieved sensitivity of 0.741, specificity of 0.755, and accuracy of 0.746.
• Alternative algorithms (k-NN, SVM, AdaBoost, deep learning) showed comparable performance.

Conclusions:

• Machine learning models integrated into the MegaTox tool can predict DILI.
• The models can identify potential DILI in early-stage clinical compounds and approved drugs.
• This approach enhances early detection of hepatotoxicity, improving drug safety.