Estimation Comparison of Pharmacokinetic Models Using MONOLIX, PKBUGS, and NONMEM
Xiaoyun Li, Andrew Lewandowski, Huafeng Zhou, A. Lawrence Gould, Kuenhi Tsai
Merck & Co., Inc.
Objectives: FO, FOCE, and Laplacia methods in NONMEM6 [1] have been shown to have difficulty converging in certain scenarios. Two newer packages, PKBUGS 1.1 [2]and MONOLIX 2.4 and 3.1 [3], feature recent computational advances and have begun to be embraced by practitioners. Although earlier studies have shown similarities in the performance of the three software packages [4][5][6], bias and reliability have not been sufficiently tested. This analysis uses several criteria to compare the three packages using results from multiple replications of two different PK model scenarios.
Methods: The 1st scenario used a one-compartment model with first-order absorption and elimination and incorporated sparse sampling where each subject was sampled at two points drawn from a set of 12 values from 0 to 70 units. The 2nd scenario used a PK model defined by two differential equations with combined first-order and saturated elimination, with inter-subject variability incorporated for two of the population parameters. 100 datasets were generated for each scenario using R for the 1st scenario and NONMEM for the 2nd scenario. Each dataset was analyzed with all three packages. From the 100 estimates produced for each model, accuracy and precision of the parameter estimates were assessed using means, mean estimation error, root mean squared error (RMSE), and the percentage of 95% confidence intervals (CIs) which cover the ture parameter value.
Results: In the sparse sampling scenario, MONOLIX outperformed NONMEM and PKBUGS since it converged each time and produced better RMSE and CI coverage. NONMEM failed to converge in 1 of the datasets because of sensitivity to the initial values. PKBUGS took the longest time to run and unreliably estimated one of the inter-subject variability parameters, but otherwise it performed comparably to NONMEM. On the 2nd scenario, NONMEM produced estimates with the smallest bias and RMSE than MONOLIX or PKBUGS. Sereral runs of PKBUGS failed to converge, yielding biased estiamted with high RMSE, but the coverage was still better than in MONOLIX. All assessments in MONOLIX 2.4 were poor. Using version 3.1 and increasing the number of "burn-in" iterations from the default number improved the estimation, but it still performed worse than the other two packages. MONOLIX also underestimated the standard error resulting in low CI coverage.
Conclusions: MONOLIX performs well with a sparse sampling scheme when all parameters are associated with inter-subject variability. Simulation results show that caution should be taken when data are fitted in a model with inter-subject variability limited to some, but not all, parameters in MONOLIX and, to a lesser extend, PKBUGS. If initial values are not the converging issue, PKBUGS may not be as desirable a choice for complex PK applications as NONMEM unless convergence is monitored carefully.
References:
[1] Beal, S.L. and L.B. Sheiner (1992). NONMEM User's Guide. part I-VII. San Francisco: NONMEM Project Group.
[2] Lunn, D.J., J. Wakefield, A. Thomas, N. Best and S. Spiegelhalter. PKBugs User Guide. V. 1.1. Dept. Epidemiology and Public Health, Imperial College School of Medicine, London, 1999.
[3] Lavielle, M. Monolix Users Guide. V. 2.4. Laboratoire de Mathmatigues, U. Paris-Sud. Orsay France. 2008.
[4] Bauer, R.J. Guzy S., and Ng C. A survey of population analysis methods and software for complex pharmacokinetics and pharmacodynamic models with examples. AAPS Journal 2007; 9:60-83.
[5]Dufful, S., B. Kirkpatrick, M.J. Carl, B. Green and N. Holford. Analysis of population pharmacokinetic data using NONMEM and WinBUGS. J. Biopharmaceutical Stat. 2004. 15:53-73.
[6] Laveille, C., M. Lavielle, K. Chatel, and P. Jacqmin. Evaluation of the PK and PK-PD libraries of MONOLIX: A comparison with NONMEM. PAGE 17(2008) Abstr 1356.
