Protein classification based on graphical encoding and convolutional neural network

Sequence signature methods to analyze protein sequences have lately attracted the attention of bioinformaticians and computational biologists. These methods are based on alignment-free approaches using Frequency Chaos Game Representation (FCGR). Indeed, there are many successful applications of these approaches such as DNA classification, sequence comparison, and phylogeny studies. Consequently, applying such methods to explore the properties of protein sequences has opened a new research area in recent years. However, there is difficulty in dealing with proteins due to high space dimensionality and the physicochemical characteristics embedded in the 20 amino acids the protein sequence is made of. Here, we explore the effectiveness of applying different strategies for encoding amino acid (protein) sequences into images using FCGR for assigning these sequences into their corresponding families and subsequently predicting their function. Reverse translation and 20-flake representation of protein sequence FCGR strategies have been examined. A Convolutional Neural Network (CNN) has been implemented to classify the performed protein images. The results propose that the 20-flake strategy has improved the accuracy of the multi-class classification task compared to the reverse translation technique.