All RNA samples had an RNA integrity number of

at least six. Transcriptome analysis was performed following the manufacturer’s recommendation in the Affymetrix Gene Chip® Expression Analysis Technical Manual (Santa Clara, CA) as previously specified (Gebel et al., 2010). The quality of Affymetrix CEL files was checked by utilizing the R packages affy, gcrma, and affyPLM (Bolstad et al., 2005, Gautier et al., 2004 and Wu et al., this website 2005). Normalized Unscaled standard error (NUSE) box plots and relative log expression (RLE) box plots were generated to identify the potential outliers. A CEL file was identified as an outlier if the median of its NUSE was beyond 1.05 or the median of its RLE was beyond 0.1. Potential spatial artifacts on arrays were checked by plotting the image and pseudo image for all the arrays. Microarray expression values were generated from the CEL files using background correction, quantile normalization, and median polish summarization. A probe set was filtered out when the 95% quantile of the

log 2 expression value was less than 7. To extract a specific and robust gene signature which can discriminate the tumors of differently exposed mice, supervised machine-learning approaches including SAM (Tusher et al., 2001) and support vector machine (Cortes and Vapnik, 1995) were applied in a 10-fold cross-validation procedure. A preliminary pathway analysis was performed using DAVID (Huang et al., 2008 and Huang et al., 2009). For continuous data, in general the arithmetic mean and the standard PLX4032 error (SE) are given as descriptive statistics, but for chemical-analytical data describing the test atmosphere the SD was calculated. Continuous data, such as organ weights, were statistically evaluated using a 1-way analysis of variance (ANOVA) followed by a Tukey test (Zar, 1984). Statistical evaluation of all neoplastic findings was carried out using an exact Trend Test (Peto et al., 1980). All calculations were performed using the Pathdata-System statistical program (Rotkreuz, Switzerland). old Non-neoplastic findings were statistically analyzed with

a non-survival adjusted Trend Test (Armitage, 1955). For all neoplastic (except lung tumors) and non-neoplastic findings a one-sided Fisher’s exact test (pairwise comparisons of control groups vs. concentration groups) was performed. For the lung tumor incidence, the Fisher Exact Test was applied for overall analysis followed by pairwise comparison. For the lung tumor multiplicity, the 1-way ANOVA was applied followed by pairwise comparison using the Tukey test (Zar, 1984). For the calculation of the discriminatory power (β = 0.2) of various study designs, the minimal detectable difference (MDD) was calculated by comparing the slopes of two linear regression lines. A t-distributed test statistic was calculated ( Sachs, 1978) by using the same variance components of actual study data for both regression lines.