Diffusion models for virtual populations and pharmacometric simulations

Clinical Snapshot

Key Findings

• P Value: Not reported; study uses distributional similarity metrics rather than hypothesis testing framework

• Effect Size: DDPM vs TVAE: categorical MAE 0.31% vs 1.07% (3.5-fold improvement); bivariate correlation MAE 0.033 vs 0.091 (2.8-fold improvement). TDM imputation MAE: Day 1 = 0.43%, Cmax Dose-9 = 0.25%, terminal phase = 0.76%

• Primary Outcome: DDPM achieved high-fidelity generation of 31-variable PDODD covariate distributions from NHANES data; TDM accurately imputed masked nivolumab PK time points across early, peak, and terminal phases

• Nnt Or Sensitivity: Not applicable (methodological modelling study). Key performance metrics: DDPM continuous KSD = 0.014 (excellent fit; KSD <0.05 threshold); SDM nivolumab KSD = 0.047 (acceptable fit); TDM imputation Pearson correlations ≥0.999 across all masked time points

• Confidence Interval: Not reported for any metric

Clinical Application

Implementation requires substantial computational infrastructure and deep learning expertise not routinely available in standard clinical pharmacology or hospital pharmacy settings. Integration into existing pharmacometric workflows (NONMEM, Monolix, R/Stan) would require significant software development. The approach is currently most feasible in academic or large pharmaceutical industry research environments. In the Australian context, virtual population simulation is relevant to TGA regulatory submissions for population PK studies and to PBS economic modelling requiring patient-level simulation. The NHANES-derived PDODD covariate model is not directly applicable to Australian patients without retraining on Australian population data (e.g., AusDiab, AIHW National Health Survey). Nivolumab (Opdivo) is TGA-approved and PBS-listed for multiple indications; improved PK simulation tools could support dose optimisation research in Australian oncology populations, particularly given differences in body composition and renal function distributions. RACGP and SHPA guidelines do not currently address generative AI-based pharmacometric simulation, and no Australian regulatory framework for synthetic patient data in drug submissions yet exists. Pharmacometricians, clinical pharmacologists, and drug developers seeking to generate synthetic virtual patient populations for population PK/PD modelling, clinical trial simulation, and regulatory submissions. Most immediately applicable to oncology PK modelling given the nivolumab dataset used.

Abstract

To evaluate diffusion model-based artificial intelligence approaches for generating virtual populations with physiological determinants of drug dosing (PDODD) and pharmacokinetic (PK) profiles. A denoising diffusion probabilistic model (DDPM) was applied to a 31-variable dataset of PDODD covariates (18 continuous, 13 binary) from the National Health and Nutrition Examination Survey and compared to a tabular variational autoencoder (TVAE). For nivolumab PK data (12,000 patients, 13 time points, 5 covariates), sequence-based diffusion model (SDM) and a time-aware diffusion model (TDM) with temporal self-attention were evaluated. The predictive performance of the TDM was evaluated by imputing masked time points. All models were trained and tested on 80%:20% partitions of the data using univariate, bivariate, and multivariate distributional similarity metrics. The diffusion model satisfactorily approximated the univariate distributions of continuous PDODD biomarkers (mean Kolmogorov-Smirnov D-statistic, KSD = 0.014), disease status frequencies (mean absolute error, MAE = 0.31%), and preserved bivariate correlations (MAE = 0.033). DDPM outperformed TVAE for categorical variables (0.31% vs. 1.07% MAE) and correlation (0.033 vs. 0.091 MAE). For nivolumab PK, SDM has KSD of 0.047 and a relative error of 1.36%. TDM accurately imputed missing PK timepoints (KSD = 0.014), reconstructing masked Day 1, Peak concentration (Cmax) Dose-9, and Terminal phase concentrations with MAE of 0.43%, 0.25%, and 0.76%, and correlations ≥ 0.999. Diffusion models demonstrated strong performance in generating cross-sectional PK covariate data and longitudinal PK profiles, capturing complex distributional and temporal dependencies. Diffusion-based approaches provide a flexible and robust framework for virtual simulations in pharmacometrics.

References

1. 1.Gadgil, P. K., Poojari, S. M., & Ramanathan, M. (2026). Diffusion models for virtual populations and pharmacometric simulations. Journal of Pharmacokinetics and Pharmacodynamics. https://doi.org/10.1007/s10928-024-09935-6