M. YAGOUBI Rached

The knowledge of the protein structural class is one of the most important sources of information in many biological fields, such as function analysis, protein structure, drug design, and protein folding. However, the protein structural class prediction is still a challenge when dealing with low similarity sequences. Therefore, the accuracy of the top-performing prediction methods remains unsatisfying, especially for proteins from the + ß class. This paper proposes a novel approach for Protein Structural Class Prediction using a Convolutional Neural Network (PSCP-CNN). Our approach consists of two stages. The first is the preprocessing stage which allows the preparation of the data. The second stage is a CNN classifier that automatically extracts the needed features for the classification. To evaluate the performance of our approach, we performed the jackknife test on four low similarity benchmark datasets: 25PDB, 640, 1189, and FC699. The experimental results show that PSCP-CNN achieved high prediction accuracy, where the overall accuracy on datasets 25PDB, 640, 1189, and FC699 is 93.8%, 94.5%, 94.0%, and 98.0%, respectively. Furthermore, comparing the results obtained with existing methods shows that PSCP-CNN outperforms state-of-the-art techniques and confirms that using a convolutional neural network allows a better prediction of protein structural classes.

The knowledge of the protein structural class is one of the most important sources of information related to protein structure or that about function analysis and drug design. But researchers still face difficulties to predict the protein structural class when it is a question about low-similarity sequences. In this paper, we propose to make the prediction using a hybrid deep learning method and handcrafted features instead of shallow classifier. We input only nine features, mostly from predicted secondary structure information, to a feed-forward deep neural network. The latter will automatically explore and extend those features through many layers and discover the representations needed for classification. The obtained results, when applying the jackknife test on two low-similarity benchmark datasets (25PDB and FC699), proved to be very significant. After comparing our method to others, it has turned out that using deep learning methods affords satisfactory performance in the field of protein structural class prediction.