Toxicogenomic dose-response models for DNA chips data from rats treated by flutamide
Colomban O. (1)(2), Naudet B. (3), Maucort-Boulch D. (3), Roy P. (3), Girard P. (2)
(1) Université de Lyon, Lyon, France; (2) EA3738 CTO, Faculté de Médecine Lyon-Sud, Université Lyon 1, Oullins, France; (3) Service de Biostatistique des Hospices Civils de Lyon, Centre Hospitalier Lyon Sud, Pierre Bénite, France.
Objectives: To fully characterize testicular toxicity in adult Wistar rats induced by flutamide (FLU), a potent antiandrogen, and to estimate the benchmark doses (BMD) modifying gene expression . To achieve this objective, changes in toxicogenomic responses (gene expression profile) in the testes, whole or fractionated, will be investigated in rats exposed to FLU at different dose levels by oral gavage for 28 consecutive days.
Methods: 42 rats were randomized between 5 different arms: vehicle (control group), 0.2, 1, 6, and 30 mg/kg body weight/day. All rats were exposed to only one dose level. The dose levels were set after taking into account previously published toxicity data generated for FLU on rats exposed for 28 days . For all 43,000 genes tested by DNA chips, the same sequence of hierarchical decision tree was applied to identify potential dose-gene-expression relationship. First, a simple linear model was applied to detect which log of gene-expression were significantly changed from simple baseline. False discovery rate was controlled during this first step. Second, for all significant changes, we applied various non-linear models, assuming homoscedasticity of residual variability: stimulation or inhibition models with exponential, Emax or logistic shape . When change from baseline was found significant, choice between linear and one of the non-linear models was performed using the Schwarz criterion (BIC). The next step was to estimate the BMD. It is usually based on the assumption of homoscedasticity and normality of residuals. Consequently normality of residual distributions was tested using Kolmogorov-Smirnof or Shapiro Wilk test. When normality was not rejected the BMD was the dose level leading to a change in predicted baseline plus k SD, with k >=1 . When residual normality was rejected, a bootstrap with 1,000 replications was used to estimate the corresponding quantile of the distribution.
Results: A significant linear change from baseline was detected for one sixth of the genes: For 52%, a stimulation model was chosen and for 48% an inhibitory model. Linear and Emax models were the main preferred significant models, while exponential and logistic models were marginal. For most models (>96%), assumption of normality of residuals was not rejected and benchmark dose was estimated as the dose leading to a difference from baseline.
Conclusions: An algorithm has been proposed to model linear and non-linear dose-effect relationship toxicity expressed by DNA chips. The algorithm allows to easily characterizing the benchmark doses for a large set of genes.
 Russell S. Toxicological Sciences 98(1), 240-248 (2007)
 Toyoda K. Arch Toxicol. 74(3) :127-132
 Benchmark Dose Software (BMDS) Version 2.1, User's Manual Version 2.0, April 30, 2009