An implementation of a hybrid method based on machine learning to identify biomarkers in the Covid-19 diagnosis using DNA sequences.

Although some people do not have any chronic disease or are not in the risky age group for Covid-19, they are more vulnerable to the coronavirus. As the reason for this situation, some experts focus on the immune system of the person, while others think that the genetic history of patients may play a role. It is critical to detect corona from DNA signals as early as possible to determine the relationship between Covid-19 and genes. Thus, the effect on the severe course of the disease of variations in the genes associated with the corona disease will be revealed. In this study, a novel intelligent computer approach is proposed to identify coronavirus from nucleotide signals for the first time. The proposed method presents a multilayered feature extraction structure to extract the most effective features using an Entropy-based mapping technique, Discrete Wavelet Transform (DWT), statistical feature extractor, and Singular Value Decomposition (SVD), together. Then 94 distinctive features are selected by the ReliefF technique. Support vector machine (SVM) and k nearest neighborhood (k-NN) are chosen as classifiers. The method achieved the highest classification accuracy rate of 98.84% with an SVM classifier to detect Covid-19 from DNA signals. The proposed method is ready to be tested with a different database in the diagnosis of Covid-19 using RNA or other signals.

Introduction

Coronavirus was first seen in the Wuhan region of China at the beginning of December 2019 [1]. It is an infectious virus that causes respiratory infections and passes from person to person. The official name of the virus has been identified by the World Health Organization (WHO) as Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) [2,3]. Although SARS-CoV-2 is relatively similar to SARS-CoV, it has some important changes in the amino acid sequence that SARS-CoV-2 aims to characterize its effects on functionality or pathogenesis [4]. On January 30, 2020, Covid-19 was declared a global health emergency by the WHO. On March 11, 2020, the virus was declared a global epidemic. Coronaviruses infect humans because of the variable number of open reading frames (ORFs) and the unique structures of spike proteins. Sequence analysis showed that the SARS-CoV-2 genome can be divided into several ORFs such as ORF1a, ORF1b, ORF3a, ORF6, ORF7a, ORF7b, and ORF8 [4]. The RNA genome has a helical tubular structure as it is surrounded by the N protein. This helix is surrounded by the nucleocapsid E protein, which in turn is associated with other structural proteins such as the M and S proteins. Surface glycoproteins on the SARS-CoV spike protein play an important role in binding to the host receptor [5]. While some people infect with Coronavirus disease(Covid-19) and get really sick, some people have only mild symptoms. In severe survivors, SARS-CoV-2 infects the lower respiratory tract, rapidly transforming it into acute respiratory distress syndrome (ARDS) requiring mechanical oxygen support [4]. Several chronic health conditions such as diabetes, hypertension, and cardiac insufficiency can play a role, and also genes can influence how the bodies of people react to viruses [6,7]. Infection is an interaction between the human, the microbe, and the environment [8,9]. Increased social distance, antibody testing for infections, and contact tracing to identify and isolate infected patients, except for the coronavirus vaccine, no effective method has been found that can prevent the spread of the epidemic and completely eradicate the virus [4]. Containing a single-stranded RNA genome ranging from 25 to 32 bases, coronavirus is divided into four main genera, alpha, beta, gamma, and delta [10,11]. Genomic variations affect complications associated with infections. Genomic factors can increase resistance or susceptibility to Covid-19 complications [[12], [13], [14]]. In studies on pandemics, it is very important to identify genomic factors that affect Covid-19 because the obtained findings can be used to cure patients or take pre-disease precautions [15,16].

Main contributions and motivation

In literature, there are many detection methods such as Polymerase Chain Reaction (PCR) based methods, microarray-based methods, and isothermal-based methods to predict Covid-19 from DNA sequences. However, these methods are based on the laboratory environment and require a lot of time and cost. The most important point that distinguishes this study from others is that it offers a low-cost machine learning-based approach to detect Covid-19 from DNA sequences, without needing a laboratory environment. In literature, there is no study to predict Covid19 infection from DNA nucleotide signals. This is the first study to create spectrogram images from DNA nucleotide signals and to extract features from these images using SVD and DWT methods. This study could encourage the researchers to use the proposed method for different datasets. The main contributions of this paper could be summarized as follows:

A different feature extraction structure including methods such as DWT, SVD, and statistical features were used in this study.

A hybrid model based on a statistical feature extractor and ReliefF algorithm was used for feature selection.

The relationship between Covid-19 and genes will be determined and will contribute to scientists about the effect of genes on the severity of the coronavirus disease.

The proposed approach obtained a classification accuracy rate of 98.84% with a SVM classifier for Covid-19 infection detection by using nucleotide signals.

Literature review

Recently, ongoing efforts have been made to develop novel Covid-19 recognition approaches using Polymerase chain reaction (PCR)-based methods [16], isothermal nucleic acid amplification-based methods, microarray-based methods [17], and newly developed methods [18]. Various modelling and forecasting approaches are introduced to understand and manage the corona epidemic [19]. In these approaches, most researchers used machine learning methods to classify computed tomography (CT) and x-ray images as a healthy or Covid-19 patient [[20], [21], [22]]. Recently, artificial intelligence (AI) and deep learning approaches have been widely used for the detection of corona. In Ref. [23], authors propose an AI-based structure for the diagnosis of Covid-19 using x-ray images. Their model classified normal and Covid cases with an accuracy performance of 98.3%. In Ref. [24], the authors classify a data set containing CT images of COVID-19 patients using a transfer learning model. The accuracy performance of their model is 79.3%. In Ref. [25], authors used the Convolutional Neural Network models to classify x-ray images and pneumonia data. The proposed approach was compared with the existing CNN models. The classification performance of their model is 89.57. Xiaowei et al. proposed a deep learning system to classify x-ray images as Covid-19, pneumonia with Influenza-A and healthy. They obtain a test accuracy of 86.7% [26]. Wang and Wong proposed the COVID-Net model for identification of Covid-19 pneumonia. They obtained accuracy performance of 92.4% [27]. In Ref. [28], authors proposed a deep learning model for identification of the pneumonia from x-ray images data. They obtained classification performance of 93.73%. In Ref. [29], authors used transfer learning models to detect the pneumonia images as normal, virus pneumonia and bacterial pneumonia. They achieved a classification accuracy of 96.39%. Alakus et al. performed the models using deep learning to predict covid-19 infection. Models were tested with 18 laboratory findings of 600 patients. Their predictive models identified patients with covid-19 at an accuracy of 86.66% [30]. In Ref. [31], authors detected coronavirus from chest images using a model based on deep convolutional neural network. They achieved accuracy performance of 98%. Ozturk et al. proposed The DarkNet model to detect COVID-19 from raw chest X-ray images. The classification performance of binary classes was 98.08% and the performance of multi-class cases was 87.02% [32]. Zhang et al. used K-NN, SVM, Random Forest, and Decision Tree methods to identify effective qualitative biomarkers to distinguish SARS-CoV-2 infection from other infectious diseases. They think that the study will help to uncover potential pathogenic mechanisms of COVID-19 and to develop new targets for vaccine design [33]. Chen et al. applied machine learning models to identify biomarkers associated with COVID-19 infection in a publicly available transcriptomic dataset. The incremental feature selection method, which includes one of the classification algorithms, was used to extract the basic biomarkers with the created feature list. It is argued that the study is effective in increasing the efficiency and accuracy of the diagnosis of COVID-19 [34]. In their study, COVİD-19 Boruta was used in gene expression to distinguish COVİD-19 severe respiratory diseases. The IFS method was then used to determine the best feature of the six immune cell types and classifications for two classification algorithms. Quantitative rules and disease status are distinguished. They think that the study will contribute to a more in-depth investigation of studies on COVID-19 [35].

The proposed hybrid model to automated detection of Covid-19

In this section, the proposed novel hybrid method based on DWT and SVD is presented for the automated detection of Covid-19. The proposed method consists of data collection and preprocessing, creation of spectrogram images, feature extraction, feature selection, and a classification stage. In the data collection and preprocessing phase, 260 samples (SARS-CoV-2 nucleotide sequences belonging to 156 samples and 104 healthy nucleotide sequences) are collected from the NCBI Virus data hub. The nucleotide signals are converted to digital signals by the Entropy-based numerical mapping technique. The digitized DNA sequences were converted to spectrogram images to be 128 window width (Hamming windowing). When constructing spectrogram images, the window length was 94, the overlap was 42, fs sample rate was 150. The obtained spectrogram images with the size of 65 × 65 have been converted to a grayscale image.

In the creation of the spectrogram images phase, digital signals are transformed into grayscale images and are resized 65 × 65 sized images. DWT, statistical feature extractor, and SVD are used for the feature extraction. Then ReliefF feature selection method is used. The SVM and k-NN are used to classify the selected features. The flow diagram, which includes all phases of the proposed method, is shown in Fig. 1 .

Fig. 1

The flow diagram of the proposed approach.

Data collection and preprocessing

In this study, the dataset consists of the nucleotide signals of 156 covid-19 patients and the nucleotide signals of 104 healthy patients. A total of 260 sample nucleotide sequences were collected from NCBI Virus data hub [36]. NCBI Virus database is an integrative, value-added resource designed to support the retrieval, display, and analysis of a collection of virus sequences and large sequence datasets. It is also a community portal for viral sequence data from RefSeq, GenBank, and other NCBI repositories. The NCBI Virus database provides high-quality virus sequences with standardized metadata for all records from International Nucleotide Sequence Database Collaboration (INSDC) databases. Table 1 shows some of groups from dataset.

Table 1

Samples of the dataset.

Accession	Race	Accession	Race
SARS-CoV-2 Group		Healthy Group
NC_045512.2	China	NM_001300741.2	China
LC594644.1	Japan	NM_001387088.1	Japan
MW364964.1	Chile	NR_109888.1	Chile
MW482885.1	USA	AJWY01002716.1	USA
MT994989	Egypt	ABCB02000016.1	Egypt
MT994632.1	Iran	AJWY01006830.1	Iran
MT820485.1	Saudi Arabia	BA000015.5	Saudi Arabia
MT233521.1	Spain: Valencia	NG_013248.1	Spain: Valencia
MT253696.1	China: Zhejiang	NC_000002.12	China: Zhejiang
MT240479.1	Pakistan: Gilgit	NR_144759.2	Pakistan: Gilgit

All the sequences selected in the sample dataset are actually ssRNA(+) protein sequences. When these RNA sequences are downloaded in the FASTA format in the gene bank, they convert into DNA sequences. Therefore, nucleotide sequences corresponding to these RNA sequences were used in the study. FASTA format is a text-based format for representing either nucleotide sequences or peptide sequences.

Covid-19 and healthy nucleotide sequences were set to be same lengths. The length of each sequence is 7000 bp. A DNA sequence does not only consist of protein-coding gene regions. It also consists of parts such as intron region, promoter region, stem_loop, mat_peptide. However, the important part is the gene regions because diseases are detected only in these gene regions. The length of each sample sequence was chosen as 7000 bases because, for a fair comparison, all samples in the dataset were selected with the gene sequence of equal length representing only the gene regions. It is desired to select the genome of equal base length that represents only the gene region in all samples. The min equal base length in all sequences was set at 7000 bp. A DNA sequence has exon, and intron promoter regions. Protein-coding regions are called exons, and non-protein-coding regions are called introns. Only exon regions are examined for the detection of genetic diseases because genetic information and genetic codes are stored in these exon regions. Therefore, only the exon regions of the raw data are examined. These exon regions in the raw DNA signals are in certain base intervals and are small in size. The nucleotide signals were converted to digital signals by Entropy based mapping technique [37,38]. In this technique, nucleotide signals are digitized according to the repetition frequency of possible codons in the signals. Codons are hidden inside the nucleotide signals, and each codon encodes a specific amino acid. In Entropy-based technique, the most important thing is codons. The entropy of codons' distribution in a DNA sequence is computed. Each codon consists of only 3 bases. By shifting 1 unit, all codons in that DNA sequence are scanned without skipping and the repetition frequency of each codon in that sequence is calculated. So, assigning the sliding window size and stride number cannot be changed. In a DNA sequence, there are 64 possible states: GAT, GGC, CAT, etc. In Equation (1), i represents 64 codons. Xi represents one of the codons. For example, X1 represents GAT, X2 is for GGC, etc. For instance, ‘Ala’ amino acid encodes ‘GCA’, ‘GCC’, ‘GCG’, ‘GCT’ codons. ‘Val’ encodes ‘GTA’, ‘GTC’, ‘GTG', ‘GTT' codons. The used formula in the algorithm of Entropy based mapping technique is given in Equation (1). The fractional Shannon Entropy is used in this formula. represents the repetition frequency of each codon in the given DNA sequence. The alpha (α) value is calculated according to Equation (2). The pseudo code of the Entropy based numerical technique is given in Algorithm 1. MATLAB 2019a has been used to run the Algorithm 1.

The pseudo code of the Entropy numerical technique

The illustration of working principle of Entropy based numerical mapping technique is shown in Fig. 2 .

Fig. 2

The illustration of working principle of Entropy based numerical technique.

Creating spectrogram images

In order to construct spectrogram images of the digitized 156 corona sequences and 104 healthy sequences, STFT has been used in a MATLAB environment. All signals have been divided into sections of length 128, windowed with a Hamming window. When constructing spectrogram images, the window length was 94, the overlap was 42, fs sample rate was 150. The obtained spectrogram images with the size of 65 × 65 have been converted to a grayscale image. Fig. 3 shows the spectrogram images of healthy and corona signals.

Fig. 3

Spectrogram images of corona and healthy signals.

Feature extraction

DWT is a type of transform where a signal is used for time-frequency analysis [39]. The wavelets are sampled at discrete intervals in this method. An analysis filter bank consisting of a pair of low and high pass filters can analyze an image. The low-pass filter provides approximate information about the image, while the high-pass filter extracts details on the edges [40,41].

The SVD is used widely as a data reduction method in machine learning [42]. It also is used image compression, denoising data, and feature extraction. SVD is a method that keeps a matrix by dividing it into 3 parts and using these three parts to regenerate the same matrix. It calculates 3 sub-matrices in a matrix A. U is an orthogonal matrix, S is a diagonal matrix and V is an orthogonal matrix [43,44]. The proposed multilayer feature extraction framework uses DWT, SVD, and the statistical feature extraction function. 142 features were extracted from spectrogram images. The pseudo-code of the proposed multilayer DWT and SVD-based feature extraction network is given in Algorithm 2. Fig. 4 shows the detection of features in spectrogram images using Algorithm 2. Codons are formed by protein-coding genes in a DNA sequence shows different color frequency distributions in the spectrogram images of Corona-carrying DNA sequences due to the repetition frequency of the codons and base mutations. With the multilayer feature extraction framework, the features representing base mutations and the distribution of codons in the spectrogram images in Fig. 4 are detected in four stages. First, the features at the edges and corners are scanned, then the horizontal, vertical, and diagonal coefficients matrices are obtained and the statistical feature set is obtained. In Algorithm 2, Harris features are the detection of corner features in an image using the Harris–Stephens algorithm. Maximally Stable Extremal Regions (MSER) is a feature detector to extract MSER regions in an image. MSER is used as a method of blob detection in images. This technique was proposed by Matas et al. to find correspondences between image elements from two images with different viewpoints [45].

Fig. 4

The graphical explanation of the detection of features in the spectrogram images.

The pseudo code of the proposed feature extraction method based on SVD and DWT

The local ternary pattern (LTP) is widely used in image processing techniques for feature extraction. These approaches first divide the image into regions. Then they extract local features. LBP could be defined by three different circular neighborhoods. Different LBP operators are shown in Fig. 5 . A label for each pixel of the image is created by the LBP operator, and these labels consist of ones and zeros. These tags are compared to pixels in the NxN neighborhood of the central pixel. Each pixel is compared with the central pixel in the LTP. If the pixel value is equal to the center pixel or greater than the center pixel, it is labeled as 1, it is labeled as 0. An example for labeling pixels with the LBP operator is given in Fig. 6 [46].

Fig. 5

Various LBP operators.

Fig. 6

An example for labeling pixels with the LBP operator.

LTP uses each member value on a vector of signal to compare the center pixel and its neighbors. For each value of the signal, binary codes are generated as a result of comparisons between neighbors [46,47].

Mathematical equations of the local ternary pattern are given in Equation (3) where Ter (p,c,k) defines ternary function.1, p-c > k −1, p + k < k

LTP also thresholds pixels at three values, using c as the value of the center pixel, k as the threshold constant, and p as the neighboring pixel. Local ternary function generates −1, 0 and 1. After thresholding, the LTP is obtained by combining neighboring pixels. When the histogram of these ternary values is calculated, because of resulting in a wide range, the ternary pattern is split into two binary patterns. Histograms of these signals are concatenated, and the features are obtained. To extract features from these histograms, it has been used a statistical feature extraction function, which contains 15 statistical moments. These are mean, median, mode, min, max, standard error, sum, count, standard deviation, kurtosis, range, sample variance, skewness, z-scores, and p-value.

Feature selection

Relief based feature selection algorithm has been used in this study. 94 discriminative features were selected with ReliefF. This method performs the feature selection process by composing a model that depends on the proximity of an element in the data set with other elements in its own class and its distance from different classes [48,49].

Classification

In the classification phase, Support Vector Machine (SVM) and k-nearest neighbors (k-NN) methods were used. For comparing performances TP (True Positive), TN (True Negative), FN (False Negative), FP (False Positive) parameters are used.

TP gives correctly identified corona number, TN gives correctly identified healthy number, FN gives incorrectly identified corona number, FP gives incorrectly identified healthy number. Sensitivity, specificity, precision, accuracy and F1 score of the proposed method are given in Equations (4), (5), (6), (7), (8)).

The performance of the proposed DWT and SVD based model was evaluated using the 5- fold cross-validation approach. 80% of the data set, which was divided into 5 parts, was used for training and 20% was used for testing. This process was applied for each part separately.

Experimental results and discussion

Totally 260 sample nucleotide sequences were considered for the detection of Covid-19 infection.156 nucleotide sequences of Covid-19 patients and 104 nucleotide sequences of healthy samples. All analog sequences were digitized by the Entropy-based digital mapping technique. The digitized signals were converted into spectrogram images with the size of 65 × 65 and these images were converted to grayscale images. A multilayer DWT and SVD-based feature extraction approach were used for the feature extraction from spectrogram images. The features were extracted with the proposed feature extraction approach. 94 most distinctive features were selected using relief, and features were classified by SVM and k-NN methods. The experimental results of k-NN and SVM classifiers are shown in Table 2 .

Table 2

Test results of the proposed method according to classifiers.

Classifiers	Folds	Sensitivity (%)	Specificity (%)	Precision(%)	Accuracy(%)	F1 Score
SVM	Fold1	97.63	98.4	97.38	99.17	99.02
	Fold2	96.89	98.11	98.19	98.39	98.47
	Fold3	97.24	98.58	97.15	97.86	97.36
	Fold4	99.72	99.12	98.26	98.64	99.81
	Fold5	98.12	99.04	94.97	96.54	99.54
	Average	97.92	98.65	97.19	98.12	98.84
k-NN	Fold1	97.81	96.48	96.23	96.89	98.12
	Fold2	95.81	98.17	95.38	98.18	97.84
	Fold3	95.26	98.48	96.19	96.75	98.47
	Fold4	94.97	99.95	95.52	96.94	97.81
	Fold5	95.85	98.17	96.83	97.19	98.21
	Average	95.94	98.25	96.03	97.41	98.09

As seen in Table 1, the best classification performance with 98.84 ± 0.28% was obtained using the SVM classifier. SVM has a higher success rate than k-NN. The ROC curve of the proposed method with the highest performance using SVM and k-NN for the detection of Covid-19 is shown in Fig. 7 .

Fig. 7

The ROC curve of the proposed method in SVM and k-NN.

In literature, there are many detection methods for Covid-19 from DNA sequences. These methods are PCR-based methods, Isothermal nucleic acid amplification-based methods, microarray-based methods, and newly developed methods. Table 3 shows a comparison of our method with the current methods in the literature for the detection of Covid-19 from DNA sequences. The common feature of these methods used is that they are molecular biology techniques. Also, these methods require various equipment and educated analysts, high cost, and high temperatures. Moreover, these methods are only possibly accomplished by a well-established laboratory. Unlike other methods, our method does not require a laboratory environment, high cost, and high temperature. Many studies have been presented in the literature for the detection of Covid-19 using chest X-ray images. However, these methods are not for detecting coronavirus from nucleic acid sequences, but for detecting Covid-19 using CT images. The proposed study achieved the highest classification accuracy rate of 98.84%.

Table 3

A comparison of the proposed method with other methods.

Authors	Methods	Dataset	Results (Accuracy)
Uhlenhaut [14]	PCR-based method	Lung cellular DNA	–
Guo et al. [15]	Microarray-based method	19 cDNAs	100%
Mani et al. [16]	Silico bioinformatics analysis approach	Indian genome sequences	>98.3% accuracy
Aslan et al. [17]	A new method based on KNN	CpG island	98.4%
Zhang et al. [33]	Random Forest(RF), KNN,SVM, Decision Tree(DT)	PSMB8, COLCA2,	RF:89.3%
Chen et al. [34]	Random Forest(RF), KNN,SVM, Decision Tree(DT)	FAM83A, LGALS3BP,	SVM:88.5%
Li et al. [35]	Boruta feature filtering, Decision Tree, Random Forest	IRF9	KNN:83.8%
		Genes with accession number GSE161731	DT:80.8%
		Single-cell with accession number E-MTAB-10026	DT: 86.7%
			KNN: 88.2%
			SVM: 93.8%
			RF: 92.3%
			Max RF:90.9% (macro F1) on B cell
The proposed method	A hybrid algorithm	Covid-19
DNA sequences	98.84%accuracy

The main advantages of the proposed method are;

The proposed multilayered feature extraction structure extracted the most effective features because this structure utilized lightweight, cognitive and effective algorithms, which are the Entropy based mapping technique, DWT, statistical feature extractor, and SVD, together. The Entropy based technique better reflects the complex structure of the nucleotide signals and performs the digitization according to the frequency of repetition of codons.

ReliefF was used to select 94 most distinctive features.

An effective DWT and SVD-based method are presented for Covid-19 infection detection with a classification accuracy rate of 98.84% in the SVD classifier.

Conclusion

In this study, a novel approach has been proposed to identify and classify Covid-19 infection from nucleotide sequences. The proposed method consists of the Entropy based numerical technique for preprocessing, DWT, SVD, and statistical feature extraction functions together for feature extraction, ReliefF for feature selection, and SVM and k-NN for classification. This is the first study to detect Covid-19 from nucleotide signals. The proposed system is able to predict Covid-19 from DNA signals using SVM and k-NN with an accuracy of 98.84% and 98.09%, respectively. The performance of the proposed method is assessed using nucleotide signals and is ready to be tested with RNA or other signals.

Author statement

Bihter Das: Data curation, Writing- Reviewing and Editing, Conceptualization, Methodology, Software, Investigation, Validation.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.