Diffusion models for virtual populations and pharmacometric simulations

Area of Science:

• Artificial Intelligence
• Pharmacometrics
• Computational Biology

Background:

• Generating realistic virtual populations is crucial for drug development and pharmacokinetic (PK) studies.
• Existing methods may not fully capture complex physiological determinants of drug dosing (PDODD) and longitudinal PK profiles.

Purpose of the Study:

• To evaluate diffusion model-based artificial intelligence for generating virtual populations with PDODD and PK profiles.
• To compare diffusion models against traditional methods like variational autoencoders for data generation tasks.

Main Methods:

• Applied a denoising diffusion probabilistic model (DDPM) to PDODD covariates from the National Health and Nutrition Examination Survey.
• Utilized sequence-based diffusion models (SDM) and time-aware diffusion models (TDM) with temporal self-attention for nivolumab PK data.
• Evaluated model performance using distributional similarity metrics, including Kolmogorov-Smirnov D-statistic (KSD) and mean absolute error (MAE).

Main Results:

• Diffusion models accurately approximated univariate distributions of PDODD biomarkers and disease frequencies, outperforming a tabular variational autoencoder (TVAE).
• DDPM preserved bivariate correlations better than TVAE for PDODD data.
• TDM demonstrated high accuracy in imputing masked PK time points, reconstructing key PK parameters with high correlation.

Conclusions:

• Diffusion models show strong performance in generating cross-sectional covariate data and longitudinal PK profiles.
• These AI approaches effectively capture complex distributional and temporal dependencies in PK data.
• Diffusion-based methods offer a flexible and robust framework for virtual simulations in pharmacometrics.