A Novel Approach to Protein Folding Prediction based on Long Short-Term Memory Networks: A Preliminary Investigation and Analysis

The Protein Folding Problem (PFP) is considered one of the most important open challenges in Biology and Bioinformatics. Long Short-Term Memory (LSTM) methods have risen recently, achieving the state-of-art performance for several Bioinformatics problems such as, protein secondary and tertiary protein structure prediction. This paper describes the application of a novel approach based on the LSTM networks to the PFP using a coarse-grained model of proteins. An specific encoding scheme for representing protein folding states is also presented. The proposed approach was evaluated by means of several experiments with a dataset of protein folding, which was obtained by Molecular Dynamics simulations. We also propose a novel method for evaluating the performance of the approach based on measures used in Bioinformatics. Furthermore, a new analysis method for protein folding pathways is presented. Results suggest that the proposed approach is able to learn the protein fold transitions. Also, it is promising for the research areas related to Bioinformatics and Computational Intelligence.