Julie Bertrand (1), Maria de Iorio (2), Celine Verstuyft (3), Monidarin Chou (4), Anne-Marie Taburet (5), David Haas (6) and David J Balding (1).
(1) University College London Genetics Institute, London, UK, (2) University College London Statistical Science Department, London, UK, (3) Departments of Molecular Genetics, Pharmacogenetic, and Hormonology, Assistance Publique Hopitaux de Paris, Hopital Bicetre, Paris, France, and Univ Paris-Sud, EA2706, France, (4) Rodolphe Merieux Laboratory, Faculty of Pharmacy, University of Health Sciences, Phnom Penh, Cambodia, (5) Clinical Pharmacy Department, Assistance Publique Hopitaux de Paris, Hopital Bicetre, Paris, France, (6) Center for Human Genetics Research, Departments of Medicine and Pharmacology, Pathology, Microbiology & Immunology, Vanderbilt University School of Medicine, Nashville, Tennessee, USA.
Rational: In population pharmacokinetics (PK), the standard genetic association analysis is a stepwise procedure. We have shown that an integrated approach, simultaneously estimating the PK parameters and the genetic effect sizes, is as powerful in large samples and more powerful to detect single nucleotide polymorphism (SNP) associations with multiple PK parameters [1]. Bayesian Variable Selection (BVS) is growing in importance in high-throughput genetic association studies, and represents another attractive alternative which can manage both complex study designs and missing genetic data.
Objectives: i) to compare in a simulation study BVS with a stepwise procedure and an integrated approach and ii) to apply BVS to the analysis of nevirapine pharmacogenetics on data from the PECAN ANRS 12154 study.
Methods: BVS was implemented, using the r2jags package, as described by Kuo and Mallik [2]: an indicator variable dictates which SNPs are associated with the PK parameter of interest. We used Bernoulli and Normal priors for the SNPs indicators and effect sizes respectively. SNPs are selected based on the posterior probability distribution of their indicator approximated by means of the Markov chain Monte Carlo algorithm.
In the simulation study, we analysed 200 data sets of N=300 subjects with n=6 concentrations from a two compartments PK model and the genotypes for 1227 SNPs, where 6 unobserved causal variants were associated to the drug oral clearance.
In the PECAN ANRS 12154 study, we analysed 129 patients with trough concentrations on two occasions (after 18 and 36 weeks of treatment) among which 10 patients also had complete PK profiles (6 samples) and the genotypes for 134 SNPs.
Results: BVS detected 216 of the 1200 true signals in the simulation versus 367 and 340 for the stepwise procedure and the integrated approach and 21 false positives versus 34 and 13, respectively.
Using BVS, we found the clearance of nevirapine in HIV-infected Cambodians to be associated with the rs7246456 (in linkage disequilibrium with rs3745274 at r2> 0.6) and with the rs2279343. Thus, we replicated two out of the three associations found in [3], where a stepwise procedure was performed using the oral clearance empirical Bayes estimates as phenotypes.
Conclusion: Although simulation study results were not yet competitive, this was an initial attempt and we will explore further developments [2].
References:
[1] Bertrand J, De Iorio M, Balding DJ. Integrating dynamic mixed effect modelling and penalized regression to explore genetic association with pharmacokinetics. Pharmacogenetics Genom. (2015). In press.
[2] O’Hara R, Sillanpaa M. A review of Bayesian variable selection methods: what, how and which. Bayesian Analysis. (2009). 4:85-11
[3] Bertrand J, Chou M, Richardson D, Verstuyft C, Leger PD, Mentré F, Taburet AM, Haas D and the ANRS 12154 Study Group. Multiple genetic variants predict steady-state nevirapine clearance in HIV-infected Cambodians. Pharmacogenetics Genom. (2012). 22:868–876.
Reference: PAGE 24 () Abstr 3484 [www.page-meeting.org/?abstract=3484]
Poster: Methodology - Covariate/Variability Models