Population Pharmacokinetic of L-DOPA and its main metabolites: Use of the Phoenix based QRPEM algorithm
Serge Guzy(1), Nathan Teuscher(2), Micha Levi(3), Steve Gorman(4) and Frank Schneider(5)
(1) POP_PHARM, Albany, CA, USA, (2) Certara, Princeton, New Jersey, USA, (3) TEVA Pharmaceuticals, Malvern, PA, USA, (4) TEVA Pharmaceuticals, West Chester PA USA, (5) TEVA Pharmaceuticals, Berlin, Germany
Objectives: To obtain population PK characteristics of L-DOPA and its metabolites in plasma and urine using the parametric, non-linear mixed effect modelling with Phoenix NLME. To show the advantage of QRPEM algorithm compared to standard FOCE-ELS. To present a new parallelization technique using large grid computing resources to solve complex problems without time consuming runs.
Methods: Blood and urine concentration over 24 hours were determined after administration of 37.5 mg carbidopa (tablets) and 150 mg L-DOPA (oral solution) 30 min after carbidopa to 11 healthy subjects. Time profiles of L-DOPA, dopamine, DOPAC, HVA and 3-OMD in plasma and cumulative amounts in urine were modeled simultaneously including double peak, and extravascular formation of dopamine.
Optimization was achieved with a new accurate parametric EM method QRPEM in Phoenix NLME that uses low discrepancy Sobol sequences, as opposed to the stochastic Monte Carlo sampling technique.
Locally initiated model runs were sent to remote computing platforms for execution and results returned to the local application using parallelization techniques in Phoenix 8 and a 300 core SGE grid hosted on Amazon Web Services by means of CFN grid software.
Results: The model estimated 23 fixed and 23 random effects. Only QRPEM had enough driving force for optimal minimization. FOCE-ELS locked multiple times into local minimums with bad diagnostics. Optimization was performed sequentially, starting with the fit of L-DOPA and dopamine data. The corresponding clearance terms split across the different paths. This resulted in satisfactory goodness-of-fit, good concordance between observed and simulated visual predictive checks and very good individual Bayesian fits for all responses. The new technique shortened run times significantly. Precision of parameters could not be assessed because the number of fixed effect parameter estimates was larger than individuals.
Conclusions: A complex model has been developed and fit to the data using a combination of optimization, adjustment and Bayesian algorithm to achieve reasonable model fits and population predictions (VPC). This predictive model can be used for extrapolation to any dosage regimen.
The QRPEM algorithm coupled with the parallel computing capabilities of Phoenix 8 permitted efficient modeling of a complicated dataset in reasonable time. The tools will increase productivity of individual modelers and expand the number and type of models to support drug developments.