MTGGF: A Metabolism Type-Aware Graph Generative Model for Molecular Metabolite Prediction.

Area of Science:

• Computational chemistry
• Drug metabolism
• Machine learning in drug discovery

Background:

• In vivo metabolism of drugs generates metabolites, posing safety challenges in drug development.
• Experimental determination of metabolites is costly and time-consuming.
• Current computational methods, rule-based and rule-free, have limitations in predicting novel metabolic reactions and characterizing molecular structures.

Purpose of the Study:

• To propose a novel metabolism type-aware graph generative framework (MTGGF) for accurate molecular metabolite prediction.
• To address the limitations of existing rule-free methods regarding structural characterization and interpretability.
• To enhance the risk evaluation of drug metabolites in drug development.

Main Methods:

• Developed a two-stage learning process: pre-training on general chemical reactions and fine-tuning on type-specific metabolic reactions.
• Employed an elaborate graph-to-graph generative model treating molecules as bipartite graphs (atoms and bonds as vertices).
• Integrated interactive attention mechanisms for analyzing molecule-metabolite relationships.

Main Results:

• The MTGGF framework demonstrated superior performance compared to state-of-the-art methods in metabolite prediction.
• Ablation studies validated the effectiveness of the graph encoding components and type-specific fine-tuning.
• Case studies revealed metabolism-type-specific crucial substructures in approved drugs.

Conclusions:

• The MTGGF framework offers a robust and interpretable approach for predicting molecular metabolites.
• The identified metabolism-type-specific substructures can aid in predicting potential safety issues.
• This framework has the potential to significantly improve the risk assessment of drug metabolites in pharmaceutical research.