Predicting in vitro assays related to liver function using probabilistic machine learning.

Area of Science:

• Toxicology
• Computational Biology
• Pharmacology

Background:

• Machine learning (ML) is increasingly used in toxicology, but limited data necessitates quantifying the uncertainty of in silico predictions.
• Reliable decision-making in toxicological assessments requires robust methods for uncertainty quantification.

Purpose of the Study:

• To develop and evaluate a probabilistic framework for predicting in vitro liver assay outcomes using integrated data modalities.
• To quantify the uncertainty associated with these in silico predictions.
• To integrate predictions into an estimation of drug-induced liver injury (DILI) probability.

Main Methods:

• Systematic comparison of various probabilistic methods for predicting in vitro liver function assays.
• Integration of multiple data modalities: chemical descriptors, gene expression, and morphological profiles.
• Generation of new experimental data for reactive oxygen species generation and hepatocyte toxicity assays.

Main Results:

• Demonstrated the performance of different data modalities within the probabilistic framework.
• Successfully integrated the framework and in vitro assay predictions to estimate DILI probability.
• Provided novel experimental data for hepatocyte toxicity and reactive oxygen species generation.

Conclusions:

• Uncertainty quantification is crucial for reliable in silico toxicity predictions.
• The developed probabilistic framework enhances the prediction of in vitro assays and DILI risk.
• This approach can contribute to a safer drug development process and reduce animal testing.