Adrien Tessier (1,2,3), Julie Bertrand (4), Marylore Chenel (3) and Emmanuelle Comets (1,2,5)
(1) IAME, UMR 1137, INSERM, F-75018 Paris, France; (2) IAME, UMR 1137, Univ Paris Diderot, Sorbonne Paris Cité, F-75018 Paris, France; (3) Division of Clinical Pharmacokinetics and Pharmacometrics, Institut de Recherches Internationales Servier, Suresnes, France; (4) University College London, Genetics Institute, London, UK; (5) INSERM CIC 1414, Univrsité Rennes 1, Rennes, Francee
Context: Pharmacogenetics is now part of many clinical trials, in particular in population pharmacokinetic/pharmacodynamic (PK/PD) studies. Several methods are available to test the association between PK/PD and genetic covariates which deal with a large number of correlated polymorphisms, but there is no gold standard. In addition, they may be applied to different PK phenotypes, such as PK parameters estimated through noncompartmental analysis (NCA) or NonLinear Mixed Effects Model (NLMEM).
We investigated four association tests; a stepwise procedure and three penalised regressions (ridge regression [1], lasso [2], hyperlasso [3]), applied to four phenotypes; concentrations observed 24 and 192h after dose, area under the curve (AUC) estimated by NCA and PK model parameters estimated by NLMEM (CL, V2 and Q). A simulation study was used to compare the combinations of association tests and phenotypes.
Methods: Simulations used a PK model developed for a compound presenting a nonlinear bioavailability F [4] along with 176 Single Nucleotide Polymorphisms (SNP). Concentrations for 200 data sets were simulated under the null (H0) and alternative hypothesis (H1) for several scenarios inspired by real clinical trials, including limited or large number of subjects, and different structural models. All methods were set to target an empirical family wise error rate of 20%. Under H1, six SNP were drawn randomly and set to affect the elimination of the drug explaining overall 30% of its variability. The number of true and false positives (TP/FP) and the detection probability of the methods were evaluated.
Results: In presence of nonlinearity and/or variability in F, the methods were more powerful to detect a gene effect when applied to a PK model parameter than with other phenotypes. When the PK was linear without variability in F, their behaviour was similar when applied to NCA or PK model phenotypes. All methods were similar in terms of detection probability and showed a low ability to detect genetic effects when the number of subjects was small. Ridge regression had the largest probability to detect SNP, but also a higher number of FP.
Conclusion: Using PK model parameters is a more versatile approach than considering NCA phenotype, with more power to detect a genetic effect except when the PK was simple with a rich design. However, all approaches needed a large number of subjects (approx. 400) to detect a clinically relevant effect, especially with infrequent variants.
References:
[1] Cule E, Vineis P, De Iorio M. Significance testing in ridge regression for genetic data. BMC Bioinformatics 2011; 12: 372. [2] Tibshirani R. Regression Shrinkage and Selection Via the Lasso. J R Stat Soc Ser B 1994; 58: 267–88. [3] Hoggart CJ, Whittaker JC, De Iorio M, Balding DJ. Simultaneous analysis of all SNPs in genome-wide and re-sequencing association studies. PLoS Genet 2008; 4: e1000130. [4] Tessier A, Bertrand J, Fouliard S, Comets E, Chenel M. High-throughput genetic screening and pharmacokinetic population modeling in drug development. 22th PAGE meeting, Glasgow, Scotland 2013; Abstract 2836.
Reference: PAGE 23 (2014) Abstr 3111 [www.page-meeting.org/?abstract=3111]
Poster: Methodology - New Modelling Approaches