Fig. 2

Schematics of the Deep Neural Map, including preprocessing, training, and post-processing. Samples are normalized, outliers are removed, and miRNAs are filtered. Preprocessed training data is the input to a 3-layer symmetric Autoencoder (AE). Once pre-trained, the latent features of the AE are forwarded to the Self-Organizing Map (SOM), which is subsequently pre-trained. Following pre-training of the AE and SOM, joint fine-tuning is performed. Post-processing consists of identification of the attention of the AE to miRNAs through the activation gradient, and identifying samples that do not cluster with their respective class