Open Access

ARTICLE

An Innovative Bispectral Deep Learning Method for Protein Family Classification

by Isam Abu-Qasmieh, Amjed Al Fahoum^*, Hiam Alquran, Ala’a Zyout

Biomedical Systems and Informatics Engineering Department, Hijjawi Faculty for Engineering Technology, Yarmouk University, Irbid, 21163, Jordan

* Corresponding Author: Amjed Al Fahoum. Email:

Computers, Materials & Continua 2023, 75(2), 3971-3991. https://doi.org/10.32604/cmc.2023.037431

Received 03 November 2022; Accepted 02 February 2023; Issue published 31 March 2023

Abstract

Proteins are essential for many biological functions. For example, folding amino acid chains reveals their functionalities by maintaining tissue structure, physiology, and homeostasis. Note that quantifiable protein characteristics are vital for improving therapies and precision medicine. The automatic inference of a protein’s properties from its amino acid sequence is called “basic structure”. Nevertheless, it remains a critical unsolved challenge in bioinformatics, although with recent technological advances and the investigation of protein sequence data. Inferring protein function from amino acid sequences is crucial in biology. This study considers using raw sequencing to explain biological facts using a large corpus of protein sequences and the Globin-like superfamily to generate a vector representation. The power of two representations was used to identify each amino acid, and a coding technique was established for each sequence family. Subsequently, the encoded protein numerical sequences are transformed into an image using bispectral analysis to identify essential characteristics for discriminating between protein sequences and their families. A deep Convolutional Neural Network (CNN) classifies the resulting images and developed non-normalized and normalized encoding techniques. Initially, the dataset was split 70/30 for training and testing. Correspondingly, the dataset was utilized for 70% training, 15% validation, and 15% testing. The suggested methods are evaluated using accuracy, precision, and recall. The non-normalized method had 70% accuracy, 72% precision, and 71% recall. 68% accuracy, 67% precision, and 67% recall after validation. Meanwhile, the normalized approach without validation had 92.4% accuracy, 94.3% precision, and 91.1% recall. Validation showed 90% accuracy, 91.2% precision, and 89.7% recall. Note that both algorithms outperform the rest. The paper presents that bispectrum-based nonlinear analysis using deep learning models outperforms standard machine learning methods and other deep learning methods based on convolutional architecture. They offered the best inference performance as the proposed approach improves categorization and prediction. Several instances show successful multi-class prediction in molecular biology’s massive data.

---

Keywords

---