Pragmatic approaches to using computational methods to predict xenobiotic metabolism

Area of Science:

• Pharmacokinetics and Drug Metabolism
• Computational Chemistry
• Medicinal Chemistry

Background:

• Accurate prediction of drug metabolism is crucial for drug discovery and development.
• Computational models are increasingly used to predict metabolic fate and sites of metabolism.
• Existing models vary in performance and may have limitations in certain chemical spaces.

Purpose of the Study:

• To investigate the performance of selected computational models for predicting metabolites and sites of metabolism.
• To assess model performance across different datasets, including homogeneous and diverse drug collections.
• To propose pragmatic approaches for utilizing metabolism prediction software.

Main Methods:

• Evaluation of MetaPrint2D-React, Meteor, and SMARTCyp software.
• Testing algorithms on two datasets: a homogeneous set of Non-Steroidal Anti-Inflammatory Drugs (NSAIDs) and paracetamol (DS1), and a diverse set of top-selling drugs (DS2).
• Analysis of prediction accuracy and identification of compounds with unpredicted metabolism.

Main Results:

• Metabolite prediction accuracy was generally better for the diverse dataset (DS2) compared to the homogeneous dataset (DS1) across all evaluated models.
• Specific compounds were identified for which none of the tested software packages could predict metabolites.
• The findings suggest that model performance is influenced by the chemical space represented in the training data.

Conclusions:

• The study highlights the strengths and limitations of current computational metabolism prediction tools.
• Diverse datasets appear more effective for model evaluation, indicating potential biases in model development.
• Pragmatic strategies, such as using cutoff values and preselection based on likely metabolic sites, are recommended for optimizing software application.