Real data comparisons of NONMEM 7.2 estimation methods with parallel processing on target-mediated drug disposition models
Tarjinder Sahota (1), Enrica Mezzalana (1), Alienor Berges (1), Duncan Richards (2), Alex Macdonald (1), Daren Austin (1), Stefano Zamuner (1)
(1) Clinical Pharmacology Modelling and Simulation, GSK, UK. (2) Academic DPU, GSK, UK
Objectives: Target-mediated drug disposition (TMDD) models are increasingly used to describe drug-target interactions. In practice, however, their use can come with long running times and convergence problems in NONMEM. Gibiansky et al., have previously used simulated data to investigate the accuracy and parallel processing efficiency of TMDD models with NONMEM 7.2.0 estimation methods1. They found all methods except BAYES gave accurate parameter estimates, standard error estimates, and high (85-95%) parallel processing efficiency. Importance sampling (IMP) was found to be the fastest exact likelihood method in terms of run time.
Here we detail our modelling experience of clinical studies using FOCE and IMP with parallelisation on two separate compounds exhibiting TMDD. We evaluate the methods for model stability and parallel processing efficiency.
Methods: Compound 1: CPHPC, a small molecule targeting serum amyloid P component (SAP), a soluble target. CPHPC was administered to patients and plasma CPHPC and SAP sampling were collected from baseline (day ‑1) to follow up (day 28). Only total concentrations (free + bound fractions of CPHPC-SAP) in plasma were available. Limited SAP recovery information was available.
Compound 2: Otelixizumab, a monoclonal antibody which is directed against human CD3ε on T lymphocytes4. Free drug in serum and free, bound and total receptors on T cells were measured using immunoassay and flow cytometry, respectively. However ~70% of free drug concentrations were below the limitation of quantification (BLQ).
Estimation method evaluation was via comparison of OFV with exact likelihood OFV (ISAMPLE=20000), run time, minimisation success criteria and parallel efficiency. Parallelisation was performed using MPI on up to four cores.
Results: FOCE linearisation was found to heavily bias the OFV in some instances. Model convergence can also fail with FOCE when datasets are associated with limitations in sampling schedule and bioanalytical assay (e.g. measurements limited to total concentrations or significant proportions of BLQ data). In contrast, both models performed well with IMP. Diagnostic performance was good and key parameters were in line with previously published values. Moreover, IMP gave significant run time improvements over FOCE. Use of parallel processing with MPI efficiently reduced run time further to levels where model development and exploration were feasible.
Conclusions: IMP is recommended over FOCE for TMDD models.
 Gibiansky, L., et al. (2012) J. PKPD. 39 (1) 17-35
 A. Berges PAGE 2012
 T. Sahota PAGE 2012
 E. Mezzalana PAGE 2012