Automated Pharmacometric Model Development by Leveraging Low-Dimensional Neural ODEs and LASSO Regression.

Area of Science:

• Pharmacometrics
• Machine Learning
• Pharmacokinetics/Pharmacodynamics

Background:

• Current pharmacometrics (PMX) model development is manual, iterative, and resource-intensive.
• Existing automated methods still rely on iterative processes and goodness-of-fit criteria for model selection.
• Neural ordinary differential equations (NODEs) show potential for complex PK/PD dynamics but lack interpretability.

Purpose of the Study:

• To develop an automated approach for generating interpretable structural models in pharmacometrics.
• To combine the data-driven capabilities of NODEs with the interpretability of mechanistic models.
• To reduce the manual effort and time required for pharmacometric model development.

Main Methods:

• Integration of neural ordinary differential equations (NODEs) with least absolute shrinkage and selection operator (LASSO) regression.
• Leveraging LASSO's feature selection to automatically propose structural models based on NODE-learned dynamics.
• Application and validation of the NODE-LASSO approach on neonatal weight, bi-exponential PK, and warfarin PK/PD data.

Main Results:

• The automated NODE-LASSO approach successfully recovered meaningful, mechanism-based structures from data.
• Demonstrated applicability across diverse scenarios including physiological and drug-specific PK/PD.
• Significantly reduced the need for extensive iterative and manual model development.

Conclusions:

• The NODE-LASSO approach offers a resource-efficient and interpretable modeling strategy for pharmacometrics.
• This method has strong potential for application in model-informed drug development and clinical research.
• Automating the proposal of mechanistic structures enhances the utility of data-driven models in PMX.