Annotating protein secondary structure from sequence

Proteins are linear chain biomolecules composed of 20 different amino acids which form secondary structures such as helixes and loops [8]. Here, we use deep recurrent memory networks and deep convectional networks for predicting secondary structure annotations on sequences from the Uniprot database. We compare different network architectures and memory models, and discuss possible improvements. Our results on several tasks are significantly better than current methods that use SVMs and feed-forward neural networks.