Gene expression profiling in toxicogenomics is often used to find molecular signature of toxicants. The range of doses chosen in toxicogenomics studies does not always represent all the possible effects on gene expression: several doses of toxicant can lead to the same observable effect on the transcriptome. This makes the problem of dose exposure prediction difficult to address. We propose a strategy allowing to gather the doses with similar effects prior to the computing of a molecular signature. The different gathering of doses are compared with criteria based on likelihood or Monte Carlo Cross Validation. The molecular signature is then determined via a voting algorithm. Experimental results point out that the obtained classifier has better prediction performances than the classifier computed according to the original labeling.