Prediction of cytochrome P450-mediated bioactivation using machine learning models and in vitro validation

Area of Science:

• Drug Metabolism and Pharmacokinetics
• Computational Chemistry
• Toxicology

Background:

• Cytochrome P450 (P450)-mediated bioactivation generates reactive metabolites (RMs), a primary cause of drug-induced hepatotoxicity and drug failure.
• Predicting P450 bioactivation is crucial for drug development to mitigate adverse drug reactions (ADRs).

Purpose of the Study:

• To develop and validate machine learning models for predicting P450-mediated bioactivation.
• To identify key molecular descriptors contributing to bioactivation prediction.
• To assess the clinical relevance of predicted bioactivation for drug safety.

Main Methods:

• Development of predictive models using Random Forest, Random Subspace, SVM, and Naïve Bayes algorithms.
• Training and testing models on a literature-derived bioactivation dataset for Benzene ring, Nitrogen heterocycle, and Sulfur heterocycle.
• Utilizing 2D descriptors such as topological indices and Burden eigenvalues.
• Validation through external dataset prediction and in vitro IC50 shift experiments.

Main Results:

• Random Forest models demonstrated high predictive performance with AUC values of 0.949, 0.973, and 0.958 for the respective test sets.
• Topological indices, 2D autocorrelations, and Burden eigenvalues were identified as significant predictive features.
• The models successfully predicted bioactivation for drugs like selpercatinib and encorafenib, with in vitro experiments confirming bioactivation potential for encorafenib and tirbanibulin.

Conclusions:

• The developed machine learning strategy provides a reliable method for predicting P450-mediated bioactivation.
• This approach can significantly aid in early-stage drug safety assessment and the design of novel drug candidates with reduced toxicity.
• The findings support the integration of computational toxicology into drug discovery pipelines to minimize ADRs.