Welcome to the the web supplement for the manuscript: Using Multitask Learning to Predict Signaling and Regulatory Pathways in Cancer


Several computational methods have been developed to reconstruct cellular response networks. High throughput data from perturbation studies, including drug response and genetic knockouts, was shown to be especially useful for this task. However, when modeling the networks following the specific perturbation being applied researchers can only use a limited set of experimental data which often leads to overfitting and to inaccurate network reconstruction. To address these issues we developed an new multi-task learning framework to reconstruct molecular response networks. In this type of machine learning framework related perturbation studies (for example, the same drug applied to different types of cancer cells) can be combined to constrain the learning of network parameters across all cells while at the same time a unique network is learned for each cell type. We used this method and thousands of genetic knockout expression experiments to reconstruct drug response networks for a number of different types of cancer. As we show, the reconstructed networks were able to correctly identify several of the key proteins and pathways involved in cancer response as shared between the cells. In addition, the method was also able to identify a number of unique pathways for each cell. Several of the cell type specific proteins identified are supported by prior work while others are novel predictions that may help explain why some drugs work well on some types of cancer but not on others.