A Machine Learning-Empowered Quantitative Structure-Activity Relationship Model for Predicting the Plasma Half-life...

Understanding a drug's plasma half-life is essential in guiding dosage regimens and optimizing therapeutic outcomes, particularly in the early stages of drug development. By using published pharmacokinetic data from Food Animal Residue Avoidance Databank, we collected 560 data points of plasma half-lives for different drugs in dogs following intravenous administration. The dataset was then preprocessed and the mean elimination half-life for each drug was selected in the final clean dataset for model training and testing. Five types of chemical descriptors and four types of supervised machine learning (ML) algorithms were employed to build ML-empowered Quantitative Structure-Activity Relationship (QSAR) models. Model performances were assessed by determination coefficient (R²) and root mean square error values. The results showed that the Deep Neural Networks model with all-combined descriptor type had the best performance with R² = 0.80 for the fivefold cross-validation set and R² = 0.57 for the testing set. Furthermore, the applicability domains of the well-trained models are shown via Williams plots. This study reports an ML-based QSAR tool in predicting the elimination half-lives of drugs in dogs based on only chemical structures. This approach can be used to support drug development in dogs and provides a basis for potential interspecies extrapolation.