Abstract:
Proteins are large and complex molecules that play a critical role in various aspects of
life. Only 20 amino acids provide millions of proteins by combining into chains called
polypeptide chains with different types of amino acids, lengths, and folds. Therefore, they
are considered the building blocks of life. In proteomics, proteins are classified into
families to achieve many goals such as predicting functional properties of novel proteins,
discovering new drugs for new diseases, etc. As biological experiments are more
expensive to deal with a large number of new proteins, one of the main computational
approaches of protein classification is deep learning. Nowadays, with the progress of
computational techniques, deep learning plays a key role in many areas. In this paper, our
goal is to offer an improved alignment-free deep learning-based method for pattern
recognition in proteins for classification. In this research work, we were based on one of
the recent deep Learning approaches called DeepFam. We designed an improved method
using the concepts which have been used in image classification and natural language
processing. We extensively experimented with using the Clusters of orthologous Groups
(COG) and G-Protein-coupled receptor (GPCR) datasets. our method showed higher
validation accuracy than DeepFam and other methods that had been experimented using
the same data sets.
Keywords: Bioinformatics, Protein family prediction, Deep learning, Deep convolutional
neural networks
