Use of colorectal cancer subtypes identified through iterative clustering to predict response to therapy.

482 Background: Colorectal cancer (CRC) is generally stratified based on genetic and epigenetic features, such as KRAS mutation and microsatellite instability status. In order to facilitate the development of new targeted drugs and treatment regimens, it is important to redefine CRC at the molecular level by identifying subtypes that are relevant for response to targeted therapy. Methods: We applied a new unsupervised approach for iteratively stratifying tumor samples using genome-wide mRNA expression data. The resulting gene expression signatures were used to subtype CRC cell line panels and publicly available CRC tumor datasets. We employed pharmacological data on the cell line panels to link the subtypes to therapy response. Results: Starting from a gene expression dataset of 63 CRC tumor samples, we employed non-negative matrix factorization (NMF) and identified two dominant CRC subtypes. In agreement with previously published results, one of the types showed a mesenchymal and the other an epithelial-like gene expression pattern. In a second step, we applied NMF on these two dominant subtypes and further stratified them into two and three subtypes, respectively. The resulting five CRC subtypes show many differences, most notably activation of specific signaling pathways. Importantly, we recovered these five subtypes in several independent, publicly available CRC datasets. This strongly suggests that the signatures capture disease-relevant features of CRC. Furthermore, we found that the different subtypes corresponded to different cell lines in a panel of CRC cell lines. The clustered CRC cell lines displayed differential responses to a number of targeted compounds, indicating that the new CRC clusters may represent disease subtypes that of differential drug sensitivity. Conclusions: The CRC subtypes discovered using our new method offer new insights into the functional and molecular processes driving CRC. Furthermore, the evidence suggests that these subtypes may differ in activated pathway status and the response to some targeted inhibitors, indicating that targeting pathways conserved in these subtypes may provide new drug discovery opportunities.