Eric A. Sherer* (1,2), Mark Sale (1,3), Steve Manuck (4), Matt Muldoon (5), Bruce G. Pollock (6,7,8), Robert R. Bies (1)
(1) Division of Clinical Pharmacology, Indiana University School of Medicine, Indianapolis, IN, USA; (2) HSR&D, Roudebush VAMC, Indianapolis, IN, USA; (3) Next Level Solutions, LLC, Raleigh NC, USA; (4) Department of Psychology, University of Pittsburgh, Pittsburgh, PA, USA; (5) Center for Clinical Pharmacology, University of Pittsburgh, Pittsburgh, PA, USA; (6) Centre for Addition and Mental Health, University of Toronto, Toronto, Ontario, Canada; (7) Rotman Research Institute, Baycrest Hospital, Toronto, Ontario, Canada; (8) Department of Psychiatry, University of Pittsburgh, Pittsburgh, PA, USA
Objectives: The scope of most manual PK model building searches is constrained by the time and effort associated with model evaluation. In an effort to more thoroughly search the global solution space, we developed a single-objective, hybrid genetic algorithm (SOHGA) approach to PK modeling building to search the global solution space for candidate models with the lowest fitness function(1). However, a limitation to SOHGA is the ad hoc nature of the fitness function used for model comparisons. A potential solution is the use of a multi-objective genetic algorithm (MOGA) which compares candidate models along multiple dimensions. The objective of this work is to compare the fits of PK models identified using manual, SOHGA, and MOGA methods.
Methods: PK models were developed independently using manual, SOHGA, and MOGA methods for each of three compounds: intravenous citalopram(2), oral perphenazine(3), and oral ziprasidone(4). All search methods were given identical options which included ADVAN/TRANS structure, inclusion of inter-occasion variability and block structure, covariate inclusion and associated function form, and form of the residual variability. For the MOGA search, we used the non-dominated, sorted genetic algorithm(5) evaluated along four dimensions: NONMEM objective function value (OFV); number of estimated parameters; convergence, covariance step, and correlation test; and global adjusted p-value from NPDE(6). The final manual and SOHGA models were compared with the MOGA candidate with the same number of parameters that converged with the lowest OFV. Models within 10 points were considered equivalent.
Results: The MOGA models were significantly better than the manual models for citalopram and ziprasidone. For ziprasidone, the difference is likely because the absorption rate constant was fixed in the manual models due to difficulties with model convergence giving MOGA model an extra degree of freedom. For citalopram, only 1 of 5 covariate effects was shared between the models so the MOGA covariate search identified a model with a lower OFV. There were no significant differences between the OFV values of the final SOHGA candidate model and the MOGA model.
Conclusions: For three test cases, a MOGA model building identified models with equal or lower OFV versus the manual approach. The MOGA approach identified models with equal OFV versus the SOHGA approach but with the intrinsic advantage of a broader view of the candidate solution space.
References:
[1] Bies RR, Muldoon MF, Pollock BG, Manuck S, Smith G, and Sale ME. A Genetic Algorithm-Based, Hybrid Machine Learning Approach to Model Selection. Journal of Pharmacokinetics and Pharmacodynamics, 33: 195-221, 2006.
[2] Muldoon MF, Mackey RH, Korytkowski MT, Flory JD, Pollock BG, and Manuck SB. The metabolic syndrome is associated with reduced central serotonergic responsivity in healthy volunteers. The Journal of Clinical Endocrinology & Metabolism 91: 718-721, 2006.
[3] Jin Y, Pollock BG, Coley K, Miller D, Marder SR, Florian J, Schneider L, Lieberman J, Kirshner M, and Bies RR. Population pharmacokinetics of perphenazine in schizophrenian patients from CATIE: Impact of race and smoking. The Journal of Clinical Pharmacology 50: 73-80, 2010.
[4] Wessels AM, Bies RR, Pollock BG, Schneider LS, Lieberman JA, Stroup S, Li CH, Coley K, Kirshner MM, and Marder SR. Population pharmacokinetic modeling of ziprasidone in patients from the CATIE Study. The Journal of Clinical Pharmacology, 11: 1587-1591, 2011.
[5] Deb K, Pratap A, Agarwal S, and Meyarivan T. A Fast and Elitist Multi-objective Genetic Algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation, 6: 2002.
[6] Comets E, Mentré F. Using simulations-based metrics to detect model misspecifications, PAGE meeting 2010.
Reference: PAGE 21 (2012) Abstr 2454 [www.page-meeting.org/?abstract=2454]
Poster: New Modelling Approaches