This manuscript describes a combined approach of unsupervised clustering followed by supervised learning that provides an efficient classification of conditions in DNA array gene expression experiments (different cell lines including some cancer types, in the cases shown). Firstly the dimensionality of the dataset of gene expression profiles is reduced to a number of non-redundant clusters of co-expressing genes using an unsupervised clustering algorithm, the Self Organizing Tree Algorithm (SOTA), a hierarchical version of Self Organizing Maps (SOM). Then, the average values of these clusters are used for training a perceptron that produces a very efficient classification of the conditions. This way of reducing the dimensionality of the data set seems to perform better than other ones previously proposed such as principal component analysis (PCA). In addition, the weights that connect the gene clusters to the different experimental conditions can be used to assess the relative importance of the genes in the definition of these classes. Finally, Gene Ontology (GO) terms are used to infer a possible biological role for these groups of genes and to asses the validity of the classification from a biological point of view.