Classification of metabolites with kernel-partial least squares (K-PLS)

Area of Science:

• Computational chemistry
• Pharmacology
• Machine learning in drug discovery

Background:

• Predicting human drug metabolism is crucial for drug discovery.
• Databases of drug metabolism data enable computational algorithm training.
• Computational approaches can identify potential drug metabolites based on molecular structure.

Purpose of the Study:

• To apply machine learning, specifically kernel-partial least squares (K-PLS), to predict human drug metabolism.
• To build predictive models using a fraction of the MetaDrug human drug metabolism database.
• To assess the classification capability of K-PLS for various metabolism rules.

Main Methods:

• Utilized a commercially available database (MetaDrug) for human drug metabolism data.
• Employed augmented atom descriptors with a K-PLS machine learning model.
• Developed models for specific metabolism rules (e.g., N-dealkylation, O-dealkylation, hydroxylations, glucuronidation, sulfation).
• Performed internal validation using leave-one-out testing and receiver operator curve statistics.

Main Results:

• K-PLS models achieved area under the curve values ranging from 0.75 to 0.84 for tested metabolism rules.
• Models successfully predicted between 61% and 79% of active molecules in leave-one-out testing.
• Demonstrated the feasibility of classifying drug metabolite formation using K-PLS.

Conclusions:

• K-PLS and similar machine learning methods (e.g., support vector machines) show promise for predicting human drug metabolite formation.
• Larger datasets with more positive examples for rare metabolism rules can improve model accuracy.
• External validation on novel molecules is recommended for further model refinement.