Automation of Structural Pharmacokinetic Model Search in NONMEM: Evaluation with Preclinical Datasets
Jeroen Schaap (1), Stefan Verhoeven (2), Gerard Vogel (3), Martijn Rooseboom (3,4) and Rene van Schaik (2)
(1) PK-PD/M&S, Clinical Pharmacology and Kinetics, Organon N.V., The Netherlands; (2) Molecular Design and Informatics, Pharmacology Oss, Organon N.V., The Netherlands; (3) DMPK & Safety, Dept. Pharmacology Oss, Organon N.V., The Netherlands; (4) Dept. Toxicology and Drug Disposition Oss, Organon N.V., The Netherlands.
Objective: One of the factors limiting further implementation of PK-PD modeling in the pharmaceutical industry is the availability of modelers. Automation of model development is therefore an attractive proposition. An alternative to the published unsupervised global model space search by means of a hybrid genetic algorithm [1], is a staged and supervised approach. The latter approach more closely reflects manual model development, with its mixed stages of model selection. We set out to implement the first building block of this approach, structural pharmacokinetic model search, and evaluate its performance in the context of small, pharmacology-type preclinical PK datasets.
Methods: Nonlinear models were fitted with NONMEM V. Model selection was performed on the basis of a hierarchy of criteria: successful covariance step, AIC (with a tolerance of 2) and number of parameters. Selected models were manually checked on the basis of goodness-of-fit and observed/predicted versus time plots. 56 Datasets originating from preclinical research, manually selected to contain the range from easy to impossible to fit, were used as test input. The experiments typically included 1-3 routes of administration. Each non-iv route was combined with iv-data; this procedure resulted in 73 datasets. The model search on this database was executed twice, on an Itanium2 machine with the Intel fortran compiler 9.1 and on an Opteron machine with g77.
Results: Screening of initial parameters estimates appeared to be the most time-consuming part of our automated model search. A staged algorithm was developed as an efficient screen. Manual inspection of automatically selected models revealed that they mostly were acceptable with the exception of models built on datasets of poor quality. The outcomes obtained at the different hardware setups revealed that (i) 8 datasets did not run due to technicalities and that (ii) 29 exactly identical objective functions (OVFs), (iii) 23 slightly different OVFs (delta < 3.7), (iv) 6 different OVFs and (v) 7 unambiguously different OFVs (delta >= 10) models were found. In groups iv and v, all but one differences were clearly related to data issues such as inconsistent increases in iv-curves. Roughly 200 model runs were performed per dataset to choose between 24 structural models, considerably more than an experienced modeler would need.
Conclusion: Structural model search automation is feasible but human supervision remains a requirement.
References:
[1] R. R. Bies, M. F. Muldoon, B. G. Pollock, S. Manuck, G. Smith, and M. E. Sale. A genetic algorithm-based, hybrid machine learning approach to model selection. J.Pharmacokinet.Pharmacodyn. 33 (2):195-221, 2006.
