Microarrays allow monitoring of thousands of genes over time periods. However, due to the low number of time points of the gene expression series, taking the temporal dependences into account when clustering the data is an hard task. Moreover, classes very interesting for the biologist, but sparse with regard to all the other genes, can be completely omitted by the standard approaches. We propose a Bayesian approach for this problem. A mixture model is used to describe and classify the data. The parameters of this model are constrained by a prior distribution defined with a new type of model that expresses our prior knowledge. These knowledge allow to take the temporal dependences into account in natural way, as well as to express rough temporal profiles about classes of interest.