In Silico Prediction of Metabolic Epoxidation for Drug-like Molecules via Machine Learning Methods.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Pharmacology

Background:

• Epoxidation is a critical drug metabolism pathway.
• Epoxide metabolites can covalently bind to DNA and proteins, posing toxicity risks.
• Experimental determination of epoxidation is challenging due to epoxide instability.

Purpose of the Study:

• To develop and validate in silico models for predicting drug epoxidation.
• To identify specific sites within compounds susceptible to epoxidation.
• To aid in the rational design of safer drug candidates by avoiding epoxidation.

Main Methods:

• Collected and curated a dataset of 829 unique epoxidation sites from 884 manually sourced data points.
• Employed three types of molecular fingerprints (1024, 2048, 4096 bits) to represent reaction sites.
• Built and evaluated 54 classification models using six machine learning algorithms, with a random 8:2 train-test split.
• Performed feature selection and external validation on the best-performing models.

Main Results:

• The optimal model achieved high accuracy (0.873) and AUC (0.944) on the test set.
• External validation demonstrated robust performance with accuracy (0.838) and AUC (0.987).
• The models accurately predict epoxidation occurrence and pinpoint susceptible sites within molecules.

Conclusions:

• In silico models based on machine learning provide a valuable approach for predicting drug epoxidation.
• These predictive models are significant tools for drug design, enabling the avoidance of potentially toxic epoxidation pathways.
• The developed models offer high accuracy and reliability for identifying epoxidation liabilities in drug candidates.