Novel neural techinques for gene expression analysis in cancer prognosis

The aim of this research thesis is the development of new methods of learning in the field of gene expression analysis of patients with cancer. With regard to unsupervised techniques, we try to design new neural networks that perform co-groupings (biclustering) to identify significant genes in certain patients. In the field of gene analysis, a common requirement is often to group genes, depending on their expression, in different samples but also to group the samples themselves based on the expression of some genes. A result of this type can be obtained through classical clustering techniques. Nevertheless, many activation patterns are common to a group of genes only under certain specific conditions, while they behave independently under other conditions. The search for these local activation patterns is the goal of biclustering. This type of analysis, unlike others, in the literature has not yet been studied in depth and, given the complexity of the scenario, requires the definition of techniques designed ad hoc which are presented in this work. The goodness of this technique has been also tested in a different research field: the face recognition.