Extracting significant features from gene expression data is a hot subject that continues to receive great attention. Many methods have been proposed in the literature to deal with this issue, but all of these methods deal with features obtained directly from the data. Since microarray data exhibit a high degree of noise, in this paper we try to reduce the noise by using double clustering approach to identify reduced set of features capable of distinguishing between two classes. Also, we showed that the transformation of the data plays a significant role in classification. We have used two forms of data, and we have used k-means and Self organizing map for clustering. Support vector machine and binary decision trees are used for classification. As a result of the conducted experiments on AML/ALL data, we have observed that CSVM is able to correctly classify the whole training and testing data when the data is log2 transformed using only few features.