Discovering discriminative graph patterns from gene expression data

We consider the problem of mining gene expression data in order to single out interesting features characterizing healthy/unhealthy samples of an input dataset. We present an approach based on a network model of the input gene expression data, where there is a labelled graph for each sample. To the best of our knowledge, this is the first attempt to build a different graph for each sample and, then, to have a database of graphs for representing a sample set. Our main goal is that of singling out interesting differences between healthy and unhealthy samples, through the extraction of "discriminative patterns" among graphs belonging to the two different sample sets. Differently from the other approaches presented in the literature, our techniques is able to take into account important local similarities, and also collaborative effects involving interactions between multiple genes. In particular, we use edge-labelled graphs and we measure the discriminative power of a pattern based on such edge weights, which are representative of how much relevant is the co-expression between two genes.