Learning & Visualizing Genomic Signatures Of Cancer Tumors Using Deep Neural Networks

Deep learning for medical diagnosis using genomics is extremely challenging given the high dimensionality of the data and lack of sufficient patient samples. Another challenge is that deep models are conceived as black boxes without much interpretation on how these complex models make predictions. We propose a deep transfer learning framework for cancer diagnosis with the capability of learning the sequence of DNA and RNA in cancer cells and identifying genetic changes that alter cell behavior and cause uncontrollable growth and malignancy. We design a new Convolutional Neural Network architecture with capabilities of learning the genomic signatures of whole-transcriptome gene expressions collected from multiple tumor types covering multiple organ sites. We demonstrate how our trained model can function as a comprehensive multi-tissue cancer classifier by using transfer learning to build classifiers for tumors lacking sufficient human samples to be trained independently. We introduce visualization procedures to provide more biological insight on how our model is learning genomic signatures and accurately making predictions across multiple cancer tissue types.