Deriving Disease Modules from the Compressed Transcriptional Space Embedded in a Deep Autoencoder