PKPy2: A Python framework for joint population pharmacokinetic and pharmacodynamic estimation, diagnostics, and uncertainty assessment

Population pharmacokinetic (PopPK) analysis usually relies on specialized, often commercial, software. PKPy introduced a Python workflow, but it derived population parameters from separately fitted individual models, so individual estimation error entered the estimated interindividual variability. We developed PKPy2, a standalone Python package that estimates structural parameters, interindividual variability, residual error, and covariate effects jointly by marginal likelihood, with optional fixed values and bounds, numerical convergence checks, and confidence intervals. An event-record interface extends estimation to NONMEM-format dosing records, nonlinear and pharmacokinetic-pharmacodynamic (PK/PD) models, correlated and interoccasion random effects, time-varying covariates, and censored observations, with standard diagnostics and uncertainty tools. In 200 simulated one-compartment intravenous datasets, all PKPy2 fits converged, and under sparse sampling the relative root mean squared error of the interindividual variance of volume was 32.49%, compared with 51.66% for PKPy. Recovery was similar to that of nlmixr2 and saemix on the same datasets, and coverage of nominal 95% intervals ranged from 90% to 100%. Predictions, likelihoods, and diagnostics of the event-record interface agreed with rxode2, independent quadrature, and nlmixr2. In clinical theophylline and warfarin data, PKPy2 estimates differed from published expert NONMEM estimates by at most 3.37% and 4.56%. In tobramycin data, the published model had a different maximum-likelihood solution; when the expert's pharmacological judgment was entered as fixed values and bounds, PKPy2 reproduced the expert estimates within 6.0%. PKPy2 provides joint PopPK and PK/PD estimation, diagnostics, and uncertainty assessment in Python.