Objective: The objective of this work was to develop a tool to simulate concentration profiles according to a user-defined pharmacokinetic and statistical model. An autocorrelation is often found in animal's concentrations between two consecutive timepoints. We therefore simulated concentration profiles without and with different autocorrelation values and study the impact on nonmem estimations.

Methods: Two scripts were written in Splus^® including user-defined pharmacokinetic (PK) models and parameters. The PK model used in both scripts was a one-compartment open-model with linear elimination and bolus administration. Clearance (Cl) and volume (V) were simulated independently according to lognormal distributions with a CV of 20%, Clpop and Vpop being set at 0.021 L/h and 0.975 L respectively, corresponding to known parameters of compound X in rat. Intra-individual variability (IAV) was generated according to two approaches: (1) a proportional error model (CV=15%), (2) the same including an autoregressive model of order 1. The autocorrelation parameter (Rho) was varied from high (0.99) to low (0) correlation. Each set of concentrations was simulated for 54 animals with 3 animals per time and 3 times per animal in order to reproduce preclinical toxicological study. 100 seeds were used to explore the impact of the randomness. Estimation of the population parameters and variability were made using Nonmem^® version V, estimation method FO, FOCE and FOCEI. First we estimated without autocorrelation then with autocorrelation in Nonmem^®.

Results: Script 1 seems to be similar to script 2 for low values of rho. The value of rho has no impact on population parameters estimations. Parameters estimations are better with FOCEI or FOCE compared to FO. The inter-individual variability (IIV) estimated with Nonmem^® increases with value of rho. The mean value of IIV of Cl increases more than V IIV. The IAV estimated with Nonmem^® decreases with increasing values of rho. The Nonmem^® estimate of Rho is poor. As a consequence, estimates of Cl and V decrease and IAV increases when autocorrelation increases.

Conclusion: Autocorrelation observed in the same individual of standard toxicological study can be simulated by introduction of the Autoregressive error model. Whatever the autocorrelation, Nonmem^® was able to determine population parameters quite well. However,  estimates of the IIV and IAV were poor between 0.3 and 0.7 probably due to the design of the study. A value of 0.3 reproduces reasonably well the autocorrelation observed in a toxicological study.