Redefining Parameter Estimation and Covariate Selection via Variational Autoencoders: One Run Is All You Need

Generative Artificial Intelligence (AI) frameworks, such as Variational Autoencoders (VAEs), have proven powerful in learning structured representations from complex, high-dimensional data. In pharmacometrics (PMX), nonlinear mixed effects (NLME) modeling is widely used to capture inter-individual variability and link covariates to characterize parameters with the goal of informing key decisions in drug research and development. This research combines the strengths of both approaches by introducing a VAE framework specifically designed for NLME modeling. The proposed method integrates the flexibility of generative AI with the interpretability and robustness of mechanism-based PMX modeling. To advance covariate selection in PMX, we replace the Evidence Lower Bound objective in VAEs with an objective function based on the corrected Bayesian information criterion. This enables the simultaneous evaluation of all potential covariate-parameter combinations, thereby allowing for automated and joint estimation of population parameters and covariate selection within a single run. Manual selection and repeated model fitting across covariate combinations are no longer required. We demonstrate the effectiveness of this combined AI-PMX approach with two representative cases. As the first generative AI-based optimization method for NLME modeling, the VAE achieves high-quality results in a single run, outperforming traditional stepwise procedures in terms of efficiency. As such, the presented approach facilitates automated model development, advancing PMX and its applications in model-informed drug development.