In the domain of gene expression data analysis, various researchers have recently emphasized the promising application of pattern discovery techniques like association rule mining or formal concept extraction from boolean matrices that encode gene properties. To take the most from these approaches, a needed step concerns gene property encoding (e.g., over-expression) and its need for the discretization of raw gene expression data. The impact of this preprocessing step on both the quantity and the relevancy of the extracted patterns is crucial. In this paper, we study the impact of discretization parameters by a sound comparison between the dendrograms, i.e., trees that are generated by a hierarchical clustering algorithm, computed from raw expression data and from the various derived boolean matrices. Thanks to a new similarity measure and practical validation over several gene expression data sets, we propose a method that supports the choice of a discretization technique and its parameters for each specific data set.