Message Passing Neural Networks Improve Prediction of Metabolite Authenticity.

Area of Science:

• Drug Metabolism and Pharmacokinetics
• Computational Chemistry
• Systems Biology

Background:

• Cytochrome P450 enzymes are crucial for drug elimination but can produce reactive metabolites, leading to drug attrition.
• Existing models predict metabolism sites or adducts, but struggle with multi-step metabolic pathways.
• Understanding network-level drug metabolism requires integrating multiple prediction models.

Purpose of the Study:

• To develop an advanced computational tool for predicting the authenticity of Phase I drug metabolite structures.
• To integrate local network structure and edge features for improved metabolic network analysis.
• To enhance the prediction of drug metabolic pathways and identify potentially reactive metabolites.

Main Methods:

• Extended XenoNet 1.0 with a bidirectional message passing neural network.
• Utilized edge-conditioned graph convolutions and jumping knowledge for network analysis.
• Trained and validated the model on a large dataset of human Phase I metabolic reactions.

Main Results:

• Achieved high accuracy in predicting metabolite authenticity (88.5% and 87.6% AUC).
• Demonstrated robustness across networks of varying complexity and identified known bioactivation pathways.
• Successfully predicted both observed and unobserved metabolites, including d,l-methamphetamine.

Conclusions:

• The enhanced XenoNet model significantly outperforms previous methods in predicting drug metabolite structures.
• The approach aids in identifying unreported metabolites and rationalizing modifications to avoid toxic pathways.
• This tool can guide drug development by predicting and mitigating risks associated with reactive metabolites.