Uncovering Population PK Covariates from VAE-Generated Latent Spaces

TL;DR

This study tackles covariate identification in population pharmacokinetics under nonlinear relationships by introducing a data-driven VAE-LASSO framework. A Variational Autoencoder learns latent representations from 10,000 synthetic tacrolimus PK profiles, while LASSO regression maps covariates to the latent space to perform sparse, interpretable covariate selection. The approach robustly retains clinically relevant covariates (SNP, age, albumin, hemoglobin) across regularization strengths and discards uninformative ones, though linear mapping limits precise PK reconstruction due to nonlinear dynamics. Overall, the method offers a scalable, model-free pathway for covariate discovery in pharmacometrics with potential to inform precision dosing and drug development.

Abstract

Population pharmacokinetic (PopPK) modelling is a fundamental tool for understanding drug behaviour across diverse patient populations and enabling personalized dosing strategies to improve therapeutic outcomes. A key challenge in PopPK analysis lies in identifying and modelling covariates that influence drug absorption, as these relationships are often complex and nonlinear. Traditional methods may fail to capture hidden patterns within the data. In this study, we propose a data-driven, model-free framework that integrates Variational Autoencoders (VAEs) deep learning model and LASSO regression to uncover key covariates from simulated tacrolimus pharmacokinetic (PK) profiles. The VAE compresses high-dimensional PK signals into a structured latent space, achieving accurate reconstruction with a mean absolute percentage error (MAPE) of 2.26%. LASSO regression is then applied to map patient-specific covariates to the latent space, enabling sparse feature selection through L1 regularization. This approach consistently identifies clinically relevant covariates for tacrolimus including SNP, age, albumin, and hemoglobin which are retained across the tested regularization strength levels, while effectively discarding non-informative features. The proposed VAE-LASSO methodology offers a scalable, interpretable, and fully data-driven solution for covariate selection, with promising applications in drug development and precision pharmacotherapy.

Figures (3)

• Figure 1: Graphical representation of the VAE-LASSO framework. The Variational Autoencoder consists of an encoder (left) that compresses pharmacokinetic signals profiles into a latent representation parameterized by a mean ($\mu$) and standard deviation ($\sigma$). The latent variables are then sampled and passed to the decoder (right), which reconstructs the original PK profiles. A LASSO regression model (bottom) is trained to predict the latent representation using patient-specific covariates, allowing for covariate selection by shrinking irrelevant features to zero.
• Figure 2: (A) Original tacrolimus PK profiles generated from the simulated dataset. (B) Reconstructed PK profiles obtained from the Variational Autoencoder (VAE). The y-axis represents tacrolimus concentration in mg/L over time. The optimal reconstruction that closely follow the original data, demonstrates the VAE’s ability to capture the underlying PK dynamics.
• Figure 3: 3D Visualization of LASSO Regression weights wcross covariates and lambda values. The height and color intensity of the bars represent the importance of each covariate at different levels of regularization $\lambda$. Brighter and taller bars indicate covariates with stronger influence (i.e., higher regression weights), while darker, shorter bars correspond to covariates with little or no contribution.The x-axis lists the covariates, y-axis the different values of the regularization parameter $\lambda$ and the z-axis the LASSO regression weights. The table denote with markers the covariates retained at each regularization level.