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

Current pharmacometrics (PMX) model development is a manual process with iterative model building, fitting, and evaluation, which can be resource-intensive and time-consuming. Existing automated model development approaches utilize algorithms that still rely on iterative processes and perform model selection based on goodness-of-fit criteria. Recent advances in machine learning and artificial intelligence, particularly neural ordinary differential equations (NODEs), have demonstrated strong potential for characterizing complex pharmacokinetic (PK) and pharmacodynamic (PD) dynamics directly from data. However, NODEs are inherently black-box models, which limits their interpretability and their ability to provide mechanistic insights, both of which are essential in PMX. We recently presented a promising concept that proposes interpretable ODE-based structural models from NODEs but relied on manually identifying functional relationships, which can be challenging and prevents full automation. In this work, we present an automated model development approach that combines NODEs with least absolute shrinkage and selection operator (LASSO) regression to automatically propose structural models based on the dynamics learned by NODEs. The approach leverages LASSO's feature selection capability, thereby linking data-driven modeling with interpretable mechanistic structures. We demonstrate the applicability of this automated NODE-LASSO model development approach in three different scenarios: neonatal weight development, bi-exponential PK data, and warfarin PK/PD data. The results indicate that our automated NODE-LASSO model development approach can recover meaningful, mechanism-based structures while reducing the need for extensive iterative and manual model development. This highlights its potential as a resource-efficient and interpretable modeling strategy for PMX and its applications in model-informed drug development and clinical research.