Eight-State Accuracy Prediction of Protein Secondary Structure using Ensembled Model

Abstract

In order to study the three dimensional structure (3D) and functionalities of proteins, the secondary structure of proteins is imperative. To date, Convolutional Neural Network (CNN) and Long Short Term Memory network (LSTM) are two networks that have been used triumphantly in the study of protein sequences. In this paper, two models for the prediction of eight-state accuracy (Q8 accuracy) for protein secondary structure are proposed, when provided with the amino acid sequence. The first model is the Concatenated Convolutional1D model (CCN1D) wherein the Convolutional 1D networks are concatenated and this model obtains only the short range information between residues. CCN1D does not obtain complete information between the amino acids present within the protein. In order to obtain long range dependency among amino acid residues, Bidirectional LSTM (BLSTM) is used. Thus in order to get accurate results, the second model known as Convolution-BLSTM (C-BLSTM) is proposed. It is a cascaded network of CNN and BLSTM. The evolutionary profiles of proteins are used as the input to both models. The Q8 Accuracy of CCN1D and C-BLSTM is 71.73% at 0.38 Dropout rate and 72.47% at Dropout rate 0.3 respectively for the CullPDB5926 dataset.