Finding structured gene signatures

In the context of gene signature identification from microarray data, a main problem is devising statistical and visual tools to interpret and understand the biological meaning of the selected genes. Most available statistical tools for gene signature extraction typically provide unstructured list of genes and lack the capability of handling correlation among genes. Recently an algorithm for feature selection, namely elastic net, was proposed allowing to deal with correlated genes in a transparent way. In this work we exploit the form of the output given by elastic net, as used in De Mol et al. (2007), to obtain a structured gene signature where genes are disposed in block of intra-correlated genes and the blocks are ranked according to a measure of the block discriminative power. After recalling how elastic net can be used to define nested lists of increasingly intra-correlated genes, we propose an ad hoc agglomerative clustering technique able to refine such a nested output by explicitly identifying modules of correlated genes. We take advantage of such a structure to visualize the correlation patterns underlying the data. The proposed procedure is validated on both synthetic data and applied to real gene expression datasets.