Integrative signatures of signaling pathway response increase robustness and accuracy of pathway predictions

MotivationAberrant cell signaling is known to drive progression of cancer and many other diseases. The study of signaling pathways within cells is central to identifying drugs that seek to modulate these pathways. Expression of pathway genes (i.e. genes that code for pathway proteins) correlates poorly with signaling pathway activity, making prediction of signaling pathway activity changes based on transcriptional disease signatures a challenging problem. Pathway architecture and response also varies across cell lines, which reflects how drug response varies across a patient population. ResultsHere, we present a transcriptional footprinting framework for predicting changes in activity of signaling pathway by integrating transcriptional signatures of genetic perturbations of pathway genes over a diverse set of cell lines into a integrative Pathway Activity Signature (iPAS). We use an unsupervised multi-task learning approach to create pathway signatures across 12 cell lines using genetic loss of function data from the LINCS project. We also use supervised learning to construct an optimal predictor based on the ensemble of 12 cell line signatures. Our methods achieve a sizeable increase in performance, as measured by prediction of pathways targeted by LINCS chemical perturbagens. AvailabilityOpen source R package iPAS is available at https://github.com/uc-bd2k/iPAS. Contactmedvedm@ucmail.uc.edu Supplementary informationSupplementary data are available online.