Deep Learning-Based Drug Half-Life Classification to Enhance Drug Development and Pharmacokinetics.

Area of Science:

• Pharmacokinetics and Drug Development
• Computational Chemistry
• Machine Learning in Medicine

Background:

• Accurate prediction of drug half-life is crucial for optimizing drug dosage and development.
• Traditional regression models struggle with pharmacokinetic variability, limiting their clinical utility.
• A need exists for more robust and interpretable methods for drug half-life assessment.

Purpose of the Study:

• To develop and validate a classification-based approach for predicting drug half-life.
• To categorize drugs into distinct short and long half-life groups using a 12-hour threshold.
• To enhance the clinical interpretability of pharmacokinetic predictions.

Main Methods:

• Utilized a convolutional neural network (CNN), specifically AlexNet, to extract molecular features.
• Employed a neural network classifier with features extracted from molecular structures.
• Implemented a holdout validation strategy with a 70/15/15 data split for training, validation, and testing.

Main Results:

• Achieved a high F1-score of 90.9% and classification accuracy of 92.3% on the test set.
• Demonstrated strong generalization capabilities with 96.2% accuracy on validation data.
• The classification framework effectively handles pharmacokinetic variability, offering superior interpretability compared to regression methods.

Conclusions:

• Proposed an efficient classification method for drug half-life prediction, aiding formulation and dosing.
• Highlighted the advantages of classification over regression for early drug development.
• Provided a scalable and robust tool for advancing pharmacokinetic research.