Sensitivity analysis of a mixture model to determine genotype/phenotype
Matthews, I and Aarons, L
University of Manchester
Objectives: Mixture models are useful in population pharmacokinetics and pharmacodynamics to characterise underlying population distributions (2 or more subpopulations) that are not adequately explained by the evaluated covariates. When applied in NONMEM, the subpopulation, to which an individual is classified, can be determined from the maximum a posteriori Bayesian post-hoc estimates. Polymorphic metabolising enzymes, such as those from the cytochrome P450 (CYP) family, may give rise to a multimodal clearance distribution. Their involvement in the elimination of many drugs means that the use of a mixture model may be of value to categorise phenotype/genotype. Several common drugs such as warfarin, metoprolol, ibuprofen, and phenytoin are eliminated via a polymorphic enzyme pathway. The aim of this study was to determine the factors that most strongly affect the ability of NONMEM mixture models to correctly predict the genotype/phenotype of an individual.
Methods: A sensitivity analysis was undertaken using NONMEM version 6ß. A one compartment disposition model with first order input and first order elimination was used to simulate the datasets. Each dataset was simulated to have 100 patients receiving a single dose at time zero with 15 log equispaced observations over 24 hours plus one observation at 30 hours. All patients were simulated without any other covariates. The parameters varied in the simulation were; the percentage of poor metabolisers (5, 15, 25, 50 %), the typical value of clearance for the poor metabolisers (0.375, 0.75, 1.125, 1.5, 1.875 L/h) relative to extensive metabolisers (3.75 L/h) and the between subject variability (15, 30, 50 %) on the structural model parameters (ka, CL, V). The analysis was repeated 10 times. The results were expressed as the percentage of false assignments as a proportion of the number of poor metabolisers.
Results: The reliability of the mixture model to correctly classify an individual as a poor or extensive metaboliser diminishes as the variability of the structural parameters increases, as the fold difference in clearance between the two populations decreases and as the proportion of poor metabolisers decreases. These results are comparable to those of Kaila et al, 2006 [1]. An empirical equation was developed to provide a decision rule for whether a mixture model may be useful for a given drug. Warfarin falls in a region of the parameter surface that would suggest that a mixture model would be inappropriate for classifying patients on warfarin therapy.
Conclusions: The most important factor affecting correct subpopulation assignment in the NONMEM mixture model was the between subject variability on the structural parameters.
References:
[1] Kaila, N., R.J. Straka, and R.C. Brundage, Mixture Models and Subpopulation Classification: A Pharmacokinetic Simulation Study and Application to Metoprolol CYP2D6 Phenotype. J Pharmacokinet Pharmacodyn, 2006
