Predicting volume of distribution with decision tree-based regression methods using predicted tissue:plasma partition...

Area of Science:

• Pharmacokinetics
• Machine Learning
• Drug Discovery

Background:

• Volume of distribution (Vss) is a key pharmacokinetic parameter indicating drug distribution in tissues.
• Accurate Vss estimation is crucial for drug development.
• Physiologically-based pharmacokinetics often utilizes tissue:plasma partition coefficients (Kt:p).

Purpose of the Study:

• To evaluate decision tree-based regression methods for estimating human Vss.
• To assess the impact of including predicted Kt:p values as features in Vss prediction models.
• To compare the accuracy of Vss prediction using molecular descriptors alone versus combined descriptors and Kt:p values.

Main Methods:

• Employed decision tree-based regression algorithms (e.g., Bagging).
• Utilized molecular descriptors and predicted tissue:plasma partition coefficients (Kt:p) as input features.
• Performed feature selection to identify optimal Kt:p subsets for prediction.

Main Results:

• Decision tree models demonstrated reasonable accuracy, comparable to inter-species Vss extrapolation.
• Models incorporating predicted Kt:p values, particularly adipose Kt:p, showed improved Vss prediction accuracy (e.g., Bagging with adipose Kt:p achieved a mean fold error of 2.29).
• Prior feature selection was beneficial when using predicted Kt:p values.

Conclusions:

• Decision tree models offer an interpretable approach for Vss estimation in drug discovery.
• Integrating diverse data sources, including predicted Kt:p, enhances Vss prediction.
• These data mining methods are valuable for predicting Vss of new chemical compounds.