Julie Bertrand (1), Maria De Iorio (2), David J. Balding (1)
(1) Genetics Institute, University College London, London, UK, (2) Department of Statistical Science, University College London, London, UK
Context: In a previous study, we have shown that an integrated approach, simultaneously estimating the PK model parameters and the genetic size effects, was, on a rich design, just as powerful as the classical stepwise procedure for covariate building to detect the effect of 6 causal variants from a large single nucleotide polymorphisms (SNPs) array [1].
Objectives: To challenge the integrated approach and the classical stepwise procedure on i/ genetic predictors of multiple PK parameters and ii/ a challenging study design. To explore a Bayesian alternative.
Methods: We simulated a two compartments PK model, with parameter values from a real-case study. From the initial scenario with N=300 subjects/n=6 sampling times and the effect of 6 SNPs on the apparent elimination clearance (CL/F) only, we now simulated i/ one scenario with the same design but the effect of 2 pairs of correlated SNPs each affecting CL/F and the central compartment apparent volume of distribution and ii/ one scenario with the same 6 causal variants on CL/F but 700 supplementary subjects having only a trough concentration.
Our integrated approach includes a penalized regression at each iteration of the Stochastic Approximation Expectation Maximization algorithm. Two penalization are considered here; Lasso and its generalization with heavier tails, HyperLasso. The full-Bayesian alternative is implemented in R using the R2jags package. For the SNP effect size we used a double-exponential (DE) prior to mimic Lasso shrinkage. The penalty parameters are set to ensure a 20% family wise error rate using an asymptotic approximation-based formula corrected for the information in the design.
Results: When genetic predictors affect multiple PK parameters, the integrated approach is more powerful than the stepwise procedure with 637 (Lasso) and 639 (HyperLasso) versus 596 true positive (on a maximum of 800).
References:
[1] Penalized regression implementation within the SAEM algorithm to advance high-throughput personalized drug therapy. Bertrand J, Bading DJ, De Iorio M. PAGE 2013.
Reference: PAGE 23 () Abstr 3205 [www.page-meeting.org/?abstract=3205]
Poster: Methodology - Covariate/Variability Models