Supplementary Figure 5: Detailed illustration of the deep learning network architecture used for predicting secondary structure and solvent accessibility of proteins.
From: Predicting the clinical impact of human mutation with deep neural networks
The input to the model is a position-weighted matrix using conservation generated by the RaptorX software (for training on Protein Data Bank sequences) or the 99-vertebrate alignments (for training and inference on human protein sequences). The output of the second to last layer, which is 51 AAs in length, becomes the input for the deep learning network for pathogenicity classification.
