Multi-descriptor approaches to oxygen binding proteins prediction and classification using deep learning

Abstract

Oxygen binding proteins play a key role in the transport and storage of oxygen through the body's cells. However, costly and time consuming biological tests can only determine a very small portion of all proteins available. This has made computational approaches increasingly essential to help biologists. The key idea behind this work is to investigate the effect of using several descriptors and deep learning to achieve better prediction. Three kinds of descriptors to represent protein sequences are considered namely amino acid composition (AAC), dipeptide composition (DC) and conjoint triad feature (CTF). Firstly, we applied deep neural networks (DNN) to single descriptor and then to multi-descriptor by investigation of the combination of descriptors. Secondly, DNN was developed to classify the oxygen binding proteins into six different classes. The experimental results show that the proposed prediction models outperform the existing methods where we obtained a Matthews correlation coefficient (MCC) of 0.9677.