Bidirectional IOHMMs and Recurrent Neural Networks for Protein Secondary Structure Prediction

Prediction of protein secondary structure (SS) is one of the classical problems in bioinformatics that are best solved using computational prediction methods based on machine learning. Current state-of-the-art predictors are based on feedforward artificial neural networks fed by a fixed-width window of amino acids, centered on the predicted residue. Using a fixed-width small window offers the advantage of architectural simplicity and allows controlling parameter overfitting. On the other hand, relevant information is also contained in distant portions of the proteins and current methods cannot exploit this information. In this chapter, we describe two alternative architectures based on noncausal (bidirectional) dynamics. These architectures can be seen as generalizations of input-output hidden Markov models or recurrent neural networks. Unlike their conventional counterparts, their outputs depend on both upstream and downstream information. This novel algorithmic idea is a first step towards architectures capable of making predictions based on variable ranges of dependencies.