Gene Expression Pattern analysis is a crucial issue in bioinformatics and biotecnology research. A common task is classification of tissues, biological pathways or diseases based on gene expression. Several machine learning tools such as artificial neural networks (NN) can be applied to aclassify gene expression profiles. Neural networks has been successfully applied in classification of multidimensional binary profiles. In this work we extend the neural network application to multi-class gene expression classification. We compare the k-class single neural network architecture (with k output classes) against a one versus all (OVA) binary classification strategy. The OVA is a combination of single neural network with single binary output by means of dividing a k-class problem in k binary classification problem. We also address the under representation problem (when one class has less samples than the other) that could detriment the generalized performance of a classifier. It is shown that the OVA classifier improves the performance compared to a single NN strategy and when the the class representation is balanced (in term of the number of samples) it´s performance is even better.