Improving Residue-Residue Contacts Prediction from Protein Sequences Using RNN-Based LSTM Network

Accurate prediction of residue-residue contacts is of crucial importance for protein structure predictions and function studies. The advantages of coevolution-based methods to predict residue-residue contacts have been made manifest in the past decade. However, the prediction of residue-residue contacts remains a challenging task since these methods need abundant homologous protein sequences to obtain higher precision. Benefiting from the rapid development and the ever-widening use of deep learning methods, we attempted to use an intelligent method to predict residue-residue contacts at an intra-protein level. The backbone of the deep learning method is a recurrent neural network (RNN) with 5-layer long short-term memory (LSTM) cells. We describe this computational model for predicting residue-residue contacts, evaluate the method on three datasets of protein chain, and report the predictive performance in obtaining 45.72%, 40.35%, 39.06% prediction precisions on long range at cut-off value L, respectively, which shows a small improvement. In addition, we also display the effects of amino acid features involved in predicting residue-residue contacts by using three unsupervised machine learning methods. The performance of our method trained on a small dataset of protein sequences sheds light on the potential usefulness of applying recurrent neural network into residue-residue contact prediction.