Predicting Secondary Structure
Neural Networks
We implemented (much of) the artificial neural network based approach described by Rost and Sander in their paper Prediction of Protein Secondary Structure at Better than 70% Accuracy. This approach encodes 13-amino-acid-long fragments of proteins (or of protein profiles, found with a sequence profiling tool) into vectors that can be fed into two connected ANNs. The ANNs are trained to classify the central amino acid in the fragment. To predict the structure of a protein, its sequence of amino acids is divided into many 13-amino-acid-long fragments which are fed into the networks.
Our implementation is in secondary_predictor_NN.py in
the prediction directory.