Figure 1

Overall strategy. A database of (in vitro) perturbation gene expression signatures is denoised for relevance in a particular cancer type (here ER+ breast cancer) by using a large training expression set of (in vivo) tumor samples, representative of that cancer type. The denoising is done with the DART-CLQ algorithm (see Figure 2), which results in a small number of clique gene modules, from which improved estimates of perturbation activity can be derived. Using the perturbation activity score matrix one can then identify associations between perturbations and clinical outcome. The same clique gene modules allow estimation of perturbation activity scores in independent in vivo tumor sample datasets and in panels of cell lines, allowing associations with outcome to be validated, and to identify potential drug treatments that may benefit certain patient subgroups. BC, breast cancer; ER+, estrogen receptor positive; MSigDB, Molecular Signatures Database.