A Novel Multi-Stage Bispectral Deep Learning Method for Protein Family Classification

Abstract

Complex proteins are needed for many biological activities. Folding amino acid chains reveals their properties and functions. They support healthy tissue structure, physiology, and homeostasis. Precision medicine and treatments require quantitative protein identification and function. Despite technical advances and protein sequence data exploration, bioinformatics' "basic structure" problem-the automatic deduction of a protein's properties from its amino acid sequence-remains unsolved. Protein function inference from amino acid sequences is the main biological data challenge. This study analyzes whether raw sequencing can characterize biological facts. A massive corpus of protein sequences and the Globin-like superfamily's related protein families generate a solid vector representation. A coding technique for each sequence in each family was devised using two representations to identify each amino acid precisely. A bispectral analysis converts encoded protein numerical sequences into images for better protein sequence and family discrimination. Training and validation employed 70% of the dataset, while 30% was used for testing. This paper examined the performance of multistage deep learning models for differentiating between sixteen protein families after encoding and representing each encoded sequence by a higher spectral representation image (Bispectrum). Cascading minimized false positive and negative cases in all phases. The initial stage focused on two classes (six groups and ten groups). The subsequent stages focused on the few classes almost accurately separated in the first stage and decreased the overlapping cases between families that appeared in single-stage deep learning classification. The singlestage technique had 64.2% +/− 22.8% accuracy, 63.3% +/− 17.1% precision, and a 63.2% +/19.4% F1-score. The two-stage technique yielded 92.2% +/− 4.9% accuracy, 92.7% +/− 7.0% precision, and a 92.3% +/− 5.0% F1-score. This work provides balanced, reliable, and precise forecasts for all families in all measures. It ensured that the new model was resilient to family variances and provided high-scoring results.

A comprehensive framework for advanced protein classification and function prediction using synergistic approaches: Integrating bispectral analysis, machine learning, and deep learning

“…Despite these results, 16 families still required assistance [39]. Using numerically encoded bispectrum images of protein sequences and a well-designed two-stage CNN model classifier, [40] introduces a new method for identifying the 16 protein families that comprise the Globin-like superfamily. Consequently, a more efficient and accurate method of protein family identification is urgently required to resolve the limitations of existing approaches.…”

“…The capacity of InterPro to amalgamate the protein signatures originating from its constituent databases into a unified and exploratory repository is a paramount attribute. Furthermore, it has the potential to leverage the unique attributes of each database in order to construct a unified and resilient database and diagnostic tool [40].…”

“…amino acid is assigned a one-hot code, wherein the one-hot code for A is 1000000000000000, the one-hot code for C is 0100000000000000, and so forth. Given the presence of unidentified amino acids within protein sequences, it is imperative to acknowledge that, under certain circumstances, an additional unit is necessary to denote the unidentified amino acid type [40]. Consequently, the binary vector's length will extend to twenty-one, as stated in reference [27].…”

“…The classification accuracy of the utilized encoding method can be improved by normalizing its outputs. This normalization process involves utilizing the mean value and standard deviation of the encoded amino acid distribution within each family, as explained in [40].…”

“…Modifications to the bispectrum can be quantifiably observed when there are alterations in the representation that lead to different quadratic nonlinearities. More information about the utilization of bispectrum is available in [40,43]. The key notion behind SqueezeNet is to remarkably diminish the number of parameters in the network by utilizing a variety of 1x1 convolutional filters, also known as "squeeze layers," and "expand layers."…”