Motivation: High-throughput gene expression profiling yields genomic signatures that allow the prediction of clinical conditions including patient outcome. However, these signatures have limitations, such as dependency on the training set, and worse, lack of generalization. Results: We propose a novel algorithm called ITI (interactome-transcriptome integration), to extract a genomic signature predicting distant metastasis in breast cancer by superimposition of large-scale protein-protein interaction data over a compendium of several gene expression datasets. Training on two different compendia showed that the estrogen receptor-specific signatures obtained are more stable (11-35% stability), can be generalized on independent data and performs better than previously published methods (53-74% accuracy). Availability: The ITI algorithm source code from analysis are available under CeCILL from the ITI companion website: http://mv.ezproxy.com.proxy.insermbiblio.inist.fr/iti. Supplementary information: Supplementary data are available at Bioinformatics online.