A Stability-Enhanced Lasso Approach for Covariate Selection in Non-Linear Mixed Effects Model

Area of Science:

• Systems Biology
• Computational Biology
• Pharmacometrics

Background:

• Non-linear mixed effects models (NLMEMs) using ordinary differential equations (ODEs) are crucial for analyzing dynamic biological systems.
• Identifying relevant covariates in high-dimensional data for these models is a significant challenge.

Purpose of the Study:

• To develop a robust method for covariate selection in ODE-based NLMEMs.
• To improve model building in complex biological systems, particularly in high-dimensional settings.

Main Methods:

• Introduction of Lasso-SAMBA (Stochastic Approximation for Model Building Algorithm), integrating Lasso regression with stability selection.
• Iterative model construction coupling penalized regression with mechanistic model estimation via the SAEM algorithm.
• Extension of the previous SAMBA strategy with a penalized, stability-driven approach to enhance robustness and reduce false discoveries.

Main Results:

• Lasso-SAMBA demonstrated superior variable selection fidelity, False Discovery Proportion (FDR) control, and computational efficiency in simulations.
• The method outperformed conventional stepwise and Bayesian variable selection techniques.
• Application to a Varicella-Zoster virus vaccination study revealed robust associations between immune response model parameters and early transcriptomic expressions.

Conclusions:

• Lasso-SAMBA offers a powerful and robust approach for high-dimensional covariate selection in ODE-based NLMEMs.
• The method has practical utility in systems vaccinology and other fields involving complex biological modeling.
• The R package implementation facilitates the application of Lasso-SAMBA in research.