Deep Learning Approach for Secondary Structure Protein Prediction based on First Level Features Extraction using a Latent CNN Structure

Abstract

In Bioinformatics, Protein Secondary Structure Prediction (PSSP) has been considered as one of the main challenging tasks in this field. Today, secondary structure protein prediction approaches have been categorized into three groups (Neighbor-based, model-based, and meta predicator-based model). The main purpose of the model-based approaches is to detect the protein sequence-structure by utilizing machine learning techniques to train and learn a predictive model for that. In this model, different supervised learning approaches have been proposed such as neural networks, hidden Markov chain, and support vector machines have been proposed. In this paper, our proposed approach which is a Latent Deep Learning approach relies on detecting the first level features based on using Stacked Sparse Autoencoder. This approach allows us to detect new features out of the set of training data using the sparse autoencoder which will have used later as convolved filters in the Convolutional Neural Network (CNN) structure. The experimental results show that the highest accuracy of the prediction is 86.719% in the testing set of our approach when the backpropagation framework has been used to pre-trained techniques by relying on the unsupervised fashion where the whole network can be fine-tuned in a supervised learning fashion.