Improving Protein Backbone Angle Prediction Using Hidden Markov Models in Deep Learning

Protein Structure Prediction (PSP) is one of the most challenging problems in bioinformatics and biomedicine. PSP has obtained significant improvement lately. This is from the growth of the protein data bank (PDB) and the use of Deep Neural Network (DNN) models since DNNs could learn more accurate patterns from more known protein structures in the PDB. Hidden Markov Models (HMM) are a widely used method to extract underlying patterns from given data. HMM profiles of proteins have been used in existing DNN models for protein backbone angle prediction (BAP), but their full potential is yet to be exploited amid the complexities involed with those DNN models. In this paper, for BAP, we propose a simple DNN model that more effectively exploits HMM profiles as features beside other features. Our proposed method significantly outperforms existing state-of-the-art methods SAP, OPUS-TASS, and SPOT-1D, and obtains mean absolute error (MAE) values of 15.45, 18.33, 6.00, and 20.68 respectively for four types of backbone angles \(\phi \), \(\psi \), \(\theta \), and \(\tau \). The differences in MAE values for all four types of angles are between 1.15% to 1.66% compared to the best known results.