A physics-informed neural network approach for estimating population-level pharmacokinetic parameters from aggregated...

Area of Science:

• Pharmacokinetics
• Computational Biology
• Machine Learning

Background:

• Pharmacokinetic literature contains rich aggregated concentration data, but tools for extracting population-level information are limited.
• Traditional physics-informed neural networks (PINNs) primarily focus on point estimates, lacking methods for distributional analysis.
• Existing methods struggle to leverage summary statistics (mean, variance) for comprehensive pharmacokinetic parameter recovery.

Purpose of the Study:

• To introduce distributional physics-informed neural networks (D-PINNs), a novel algorithm for statistical modeling within the PINN framework.
• To enable the recovery of population-level pharmacokinetic parameter distributions from aggregated concentration means and variances.
• To demonstrate the capability of D-PINNs in analyzing both simulated and real-world pharmacokinetic data, including minimal physiologically-based pharmacokinetic (mPBPK) models.

Main Methods:

• Developed D-PINNs, integrating distributional assumptions into the PINN optimization process using neural networks to predict concentration mean and variance.
• Employed a sampling-based procedure within the residual network, utilizing ordinary differential equation (ODE) systems to compute the physics-informed loss.
• Accounted for interindividual variability via parameter distributions and measurement noise through a residual error model.

Main Results:

• D-PINNs demonstrated high accuracy in estimating parameter distributions and residual errors on simulated data from a one-compartment pharmacokinetic model.
• The framework successfully recovered population-level kinetic parameter distributions from aggregated plasma concentration data for monoclonal antibodies (mAbs) using a minimal physiologically-based pharmacokinetic (mPBPK) model.
• Benchmarking against Markov chain Monte Carlo (MCMC) confirmed the efficacy of the D-PINNs methodology.

Conclusions:

• D-PINNs offer a powerful new approach for statistical pharmacokinetic modeling, enabling the extraction of population parameter distributions from aggregated data.
• The methodology effectively handles interindividual variability and measurement noise, providing robust pharmacokinetic insights.
• This work advances the application of machine learning in pharmacokinetics, facilitating more comprehensive analysis of published literature data.