A machine learning approach to population pharmacokinetic modelling automation

Area of Science:

• Pharmacokinetics
• Computational Biology
• Drug Development

Background:

• Population pharmacokinetic (PopPK) models are essential for understanding drug disposition in diverse patient groups.
• Traditional PopPK model development is time-consuming and resource-intensive.
• An automated approach can streamline this process.

Purpose of the Study:

• To develop and validate an automated, out-of-the-box method for PopPK model development.
• To leverage pyDarwin's optimization algorithms for efficient extravascular drug model identification.
• To reduce the manual effort and time required for PopPK analysis.

Main Methods:

• A generic model search space for extravascular drugs was proposed.
• A penalty function was developed to prevent over-parameterization and ensure parameter plausibility.
• Bayesian optimization with a random forest surrogate and local search using pyDarwin was employed.

Main Results:

• The automated approach identified comparable model structures to expert-developed models in under 48 hours.
• Fewer than 2.6% of potential models were evaluated, demonstrating efficiency.
• Ablation studies confirmed the effectiveness of the penalty function and global search.

Conclusions:

• A unified penalty function and model space within pyDarwin enable automatic PopPK model structure identification for various drugs.
• This automated method simplifies PopPK analysis, making it more accessible.
• Adoption can accelerate analysis, enhance model quality, and improve reproducibility.