A Combination of Support Vector Machines and Bidirectional Recurrent Neural Networks for Protein Secondary Structure Prediction

Predicting the secondary structure of a protein is a main topic in bioinformatics. A reliable predictor is needed by threading methods to improve the prediction of tertiary structure. Moreover, the predicted secondary structure content of a protein can be used to assign the protein to a specific folding class and thus estimate its function. We discuss here the use of support vector machines (SVMs) for the prediction of secondary structure. We show the results of a comparative experiment with a previously presented work. We measure the performances of SVMs on a significant non-redundant set of proteins. We present for the first time a direct comparison between SVMs and feed forward neural netwoks (NNs) for the task of secondary structure prediction. We exploit the use of bidirectional recurrent neural networks (BRNNs) as a filtering method to refine the predictions of the SVM classifier. Finally, we introduce a simple but effective idea to enforce constraints into secondary structure prediction based on finite-state automata (FSA) and Viterbi algorithm.