Building Survival Models from Genomic Data with Google TensorFlow

Large-scale cancer genomic research, such as TCGA project generated huge amount of high-dimensional data, yet most features are under-explored for clinical application. Previously, our mentor has developed a cox neural network model that can predict patients' survival time using TCGA genomic data with extremely high accuracy. Here we propose to explore methods that rank the importance of input genomic/clinical features for this cox model based on both prediction accuracy and biological information of these features to understand the underlying biology of cancer and translate the existing implementationin Theano library to TensorFlow with comprehensive documentation and hopefully develop a user-friendly web-server for average biologist/clinicians to use and draft a manuscript.