Predicting Drug-Induced Liver Injury with Bayesian Machine Learning.

Area of Science:

• Drug Discovery and Development
• Toxicology
• Computational Biology

Background:

• Drug-induced liver injury (DILI) poses significant risks in drug development, potentially causing severe morbidity or mortality and leading to drug attrition.
• Predicting DILI is challenging due to multifactorial mechanisms and overlapping etiologies, influenced by factors like physicochemical properties, dose, and hepatic function.
• There is a critical need for improved, human-relevant predictive models to enhance hepatotoxicity risk assessment during drug discovery.

Purpose of the Study:

• To develop and validate a quantitative, mechanistic risk assessment model for predicting DILI severity during candidate nomination.
• To integrate data from in vitro hepatic safety assays with physicochemical and exposure variables using Bayesian machine learning.
• To improve the accuracy and reliability of hepatotoxicity risk prediction in early-stage drug discovery.

Main Methods:

• Utilized a diverse set of 96 compounds with varying DILI severity (no/low-, medium-, high-severity).
• Employed in vitro assays including hepatic spheroids, BSEP, mitochondrial toxicity, and bioactivation.
• Integrated assay data with physicochemical properties (cLogP) and exposure (Cmax_total) using a Bayesian machine learning model with a data-driven prior distribution.

Main Results:

• The Bayesian model accurately predicted continuous DILI severity with uncertainty.
• Achieved a balanced accuracy of 63% for predicting DILI severity categories on held-out samples.
• Demonstrated high performance for binary DILI prediction (yes/no) with 86% balanced accuracy, 87% sensitivity, and 85% specificity.

Conclusions:

• Integrating mechanistically relevant hepatic safety assays with Bayesian machine learning significantly improves DILI risk prediction.
• The developed model offers a quantitative and mechanistic approach to risk assessment, aligning with FDA recommendations.
• This approach enhances the ability to identify and mitigate hepatotoxicity risks early in drug discovery, reducing drug attrition.