Reconstruction and Visualization of Protein Structures by exploiting Bidirectional Neural Networks and Discrete Classes

Abstract

In recent years, Deep Learning techniques have achieved some success in bioinformatics tasks, including protein conformation prediction. In this work, we propose a Bidirectional Long Short-Term Memory (BLSTM) network system, called Human Proteins Angles Prediction (HPAP), in order to improve the prediction of dihedral angles of proteins. We have introduced a discrete subdivision in classes of 5° for protein torsion angles and four new features related to accessible surface area and volume. In total there are 73 classes (72 classes include the angles between -180° and 180°, a further class is used to code the free angles at the beginning of the sequence) with a maximum expected error of ±2.5°. We have tested three model variants in several parameter combinations. With our model, we have obtained a decrease of the mean absolute error of about 2° for the $\psi$ angle. Although our dataset is reduced in size, the accuracy of $\varphi$ and $\psi$ angles is comparable to the existing methods. Predicting angles accurately is useful for accurately reconstructing the three-dimensional structure of a protein. In this context, the prediction is limited to the $\varphi$ and $\psi$ angles and we will visualize what happens locally when a prediction is correct. In case the prediction is far from true angles, even a small error can deconstruct the backbone.