Engineering Science and Technology, an International Journal 24 (2021) 839–847

Contents lists available at ScienceDirect

Engineering Science and Technology,

an International Journal
journal homepage: www.elsevier.com/locate/jestch

Full Length Article

A new COVID-19 detection method from human genome sequences

using CpG island features and KNN classifier
Hilal Arslan a,⇑, Hasan Arslan b
a
Department of Computer Engineering, Izmir Bakircay University, Izmir, Turkey
b
Department of Mathematics, Erciyes University, Kayseri, Turkey

a r t i c l e i n f o a b s t r a c t

Article history: Various viral epidemics have been detected such as the severe acute respiratory syndrome coronavirus
Received 13 October 2020 and the Middle East respiratory syndrome coronavirus in the last two decades. The coronavirus disease
Revised 18 December 2020 2019 (COVID-19) is a pandemic caused by a novel betacoronavirus called severe acute respiratory syn-
Accepted 30 December 2020
drome coronavirus-2 (SARS-CoV-2). After the rapid spread of COVID-19, many researchers have investi-
Available online 9 January 2021
gated diagnosis and treatment for this terrifying disease quickly. Identifying COVID-19 from the other
types of coronaviruses is a difficult problem due to their genetic similarity. In this study, we propose a
Keywords:
new efficient COVID-19 detection method based on the K-nearest neighbors (KNN) classifier using the
COVID-19
SARS-CoV-2
complete genome sequences of human coronaviruses in the dataset recorded in 2019 Novel
K-Nearest Neighbors Coronavirus Resource. We also describe two features based on CpG island that efficiently detect
CpG islands COVID-19 cases. Thus, genome sequences including approximately 30,000 nucleotides can be repre-
Human coronaviruses sented by only two real numbers. The KNN method is a simple and effective non-parametric technique
for solving classification problems. However, performance of the KNN depends on the distance measure
used. We perform 19 distance metrics investigated in five categories to improve the performance of the
KNN algorithm. Some efficient performance parameters are computed to evaluate the proposed method.
The proposed method achieves 98.4% precision, 99.2% recall, 98.8% F-measure, and 98.4% accuracy in a
few seconds when any L1 type metric is used as a distance measure in the KNN.
Ó 2020 Karabuk University. Publishing services by Elsevier B.V. This is an open access article under the CC
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction and HKU1-CoV cause simple respiratory traces, SARS-CoV and

Coronaviruses, which are positive sense and single-stranded
RNA viruses, are known to have the largest viral genomes among
all RNA viruses [1]. The family of coronaviruses has been classified
into four genera [2]: alphacoronavirus (AlphaCoV), betacoron-
avirus (BetaCoV), gammacoronavirus (GammaCoV), and deltacoro-
navirus (DeltaCoV). AlphaCoV and BetaCoV infect mammalian
hosts, however, GammaCoV and DeltaCoV mainly infect bird spe-
cies [3].
There have been recorded various coronaviruses, which are the
types of AlphaCoV and BetaCoV. Human coronavirus 229E (229E-
CoV) and human coronavirus NL63 (NL63-CoV) are the types of
AlphaCoV. Moreover, human coronavirus HKU1 (HKU1-CoV), Sev-
ere Acute Respiratory Syndrome coronavirus (SARS-CoV), and Mid-
dle East Respiratory Syndrome coronavirus (MERS-CoV) are the
types of BetaCoV recorded recently. While 229E-CoV, NL63-CoV
⇑ Corresponding author.
E-mail addresses: hilal.arslan@bakircay.edu.tr (H. Arslan), hasanarslan@erciyes.
edu.tr (H. Arslan).
MERS-CoV cause fatal and dangerous respiratory infections.
SARS-CoV was first identified in the Guangdong province of
southern China on 16 November 2002 [4] and MERS-CoV emerged
in June 2012 in Jeddah, Saudi Arabia [5]. In late December 2019, a
novel coronavirus SARS-CoV-2 emerged in China and spread
rapidly throughout the world. The SARS-CoV-2 virus, which is
genetically similar to SARS-CoV, caused a severe illness known as
COVID-19 disease and a serious number of deaths worldwide.
Since several early infected people visited a local seafood Market
in December 2019 in Wuhan city of China, the virus is therefore
thought to be a pathogen (SARS-CoV-2 pathogen) that jumped
from an animal to a human and that caused an infectious disease.
Similarity rate between SARS-CoV-2 and a bat coronavirus is 96%
[6]. The outbreak COVID-19 was declared as a global pandemic
on 11 March 2020 by the World Health Organization (WHO). As
of 8 October 2020, the disease spread to 188 countries and territo-
ries has infected over 36.2 million people and has caused more
than 1.05 million deaths; more than 25.2 million people have
recovered from this illness.

https://doi.org/10.1016/j.jestch.2020.12.026
2215-0986/Ó 2020 Karabuk University. Publishing services by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
H. Arslan and H. Arslan
The most common symptoms of COVID-19 disease are cough,
fever, gastrointestinal and musculoskeletal symptoms as well as
loss of taste or smell. Shortness of breath and chest pressure or
pain are among less common symptoms [7]. Because these com-
mon symptoms are similar to the common flu, it is difficult to con-
duct an early diagnosis of SARS-CoV-2. It is essential to quickly
identify positive cases since it spreads rapidly and poses a threat
to the public health system. Furthermore, there is still no specific
antiviral drug recommended for the treatment of the novel coron-
avirus disease other than supportive care so far. In the period of
pre-clinical researches, some medications already prescribed for
other diseases have shown positive effects against COVID-19 virus.
Due to the absence of an effective COVID-19 treatment, early
detection of this life-threatening infectious disease plays a very
important role in all medical therapies and in the prevention of
the spread of the disease. Recently, many data scientists have been
working extensively on remarkable features of the virus. For this
purpose, artificial intelligence applications and machine learning
methods have also been used successfully to speed up the diagno-
sis process of the COVID-19 cases [8–10]. Therefore, the learning
based classification models will not only reduce the burden on
healthcare professionals but will also facilitate early diagnosis of
the disease.
In this paper, we introduce an effective classification method to
distinguish SARS-CoV-2 from common human coronaviruses,
which are AlphaCoV, BetaCoV-1, MERS-CoV, HKU1-CoV, NL63-
CoV, and 229E-CoV using CpG island features and KNN classifica-
tion method. Main contributions of this research can be summa-
rized as follows:

The choice of the differentiable features based on characteristic
of SARS-CoV-2 is a critical step to improve classification perfor-
mance. In this study, we propose CpG island features to differ-
entiate SARS-CoV-2 from the other human coronaviruses.

We propose a robust prediction method by using KNN classifier
with any L1 type distance metric selected from among 19 dis-
tance metrics investigated in five categories, which are L1 type,
L2 type, vicissitude, inner product metrics, and the other types
of metrics.

We construct a larger dataset containing almost all types of
these human coronaviruses which are genetically similar to
SARS-CoV-2. They are AlphaCoV, BetaCoV_1, MERS-CoV, NL63-
CoV, HKU1-CoV, and 229E-CoV.

The proposed method efficiently detects COVID-19 cases in a
few seconds on the relatively large dataset.
The rest of the paper is organized as follows. In Section 2, a lit-
erature survey on COVID-19 is presented. In Section 3, the K-
nearest neighbors (KNN) method and distance measures used in
the KNN method are summarized. The proposed classification
strategy to detect COVID-19 cases is introduced in Section 4. Sec-
tion 5 reports experimental results and evaluation of the experi-
mental observations conducted by various distance metrics. In
Section 6, we compare the proposed method with the other detec-
tion methods. Finally, Section 7 presents concluding remarks and
future directions.
2. Related works
Deep learning and machine learning techniques have been
widely used in a variety of research areas such as big data analysis
[11–13], image classification [14], face detection [15,16], and dis-
ease prediction [17–21]. The computer based diagnostic systems
developed by the help of artificial intelligence techniques will
speed up early diagnosis of COVID-19 and thus will decrease the
840
Engineering Science and Technology, an International Journal 24 (2021) 839–847
workload on health-workers. Chen et al. [22] published a compre-
hensive survey about studies performed by using artificial intelli-
gence in the literature on the fight against COVID-19. In addition,
they observed that machine learning, deep learning and artificial
neural networks technologies have been successfully implemented
at almost every stage of combating COVID-19 when compared to
other coronaviruses such as SARS-CoV and MERS-CoV.
Randhawa et al. [23] suggested a combination of supervised
machine learning with digital signal processing methods (ML-
DSP) for an accurate and scalable taxonomic classification of geno-
mic sequences. They matched each genomic sequence to discrete
values corresponding to its genomic signals. They employed six
supervised machine learning classifiers (linear discriminant, linear
support vector machine, quadratic support vector machine, fine
KNN, subspace discriminant, and subspace KNN) to detect SARS-
CoV-2. They used 29 SARS-CoV-2 genome sequences and 20 gen-
ome sequences for each of alphacoronavirus, betacoronavirus,
and delta-coronavirus. They observed that the linear discriminant
method achieved 100% accuracy. Naeem et al. [24] proposed a
method to distinguish COVID-19, SARS-CoV, and MERS-CoV viruses
by using the K-nearest neighbors and the trainable cascade-
forward back-propagation neural network methods. They
extracted genomic signal processing features using a dataset that
contains 76 genome sequences for each type of coronavirus from
the National Center for Biotechnology Information (NCBI). Their
results showed that performance of the KNN algorithm was higher
than the cascade-forward back-propagation neural network in all
COVID-19/SARS-CoV, COVID-19/MERS-CoV and COVID-19/SARS-
CoV/MERS-CoV classification processes, and achieved an accuracy
of 100%. Batista et al. [25] developed a new method to predict
whether patients in the emergency care unit are COVID-19 positive
or not by using five supervised machine learning algorithms, which
are neural networks, random forests, logistic regression, support
vector machines, and gradient boosting regression trees. They col-
lected data from 235 patients, and they used 15 different types of
features such as age, gender, hemoglobin etc. Their experimental
results showed that the support vector machine classifier achieved
the highest performance with AUC value of 84.7% among the other
classifiers.
Unal and Dudak [26] studied on diagnosis of COVID-19 viral dis-
ease. They applied Naive Bayes, KNN, support vector machines and
decision tree algorithms to the dataset named as COVID-19 Mexico
Patient Health Dataset. The dataset consists of 95839 cases
recorded by Mexican government, and 19 different types of fea-
tures like as the sex of the patient, age of the patient, the state of
pneumonia and intubation as well as the state of many other dis-
eases. They performed four types of supervised machine learning
algorithms. Their experimental results showed that support vector
machine achieved the best predictive performance with the classi-
fication accuracy of 100%.
Although there are only few studies detecting COVID-19 cases
from genome sequences, a number of papers have been published
for detecting COVID-19 cases from X-ray or computed tomography
(CT) images [27–34], recently. Barstugan et al. [27] developed a
COVID-19 classification method by using the support vector
machine classifier on a dataset including different types of 150
abdominal CT images. Their experimental results showed that their
proposed method achieved an accuracy of 99.6%. Ozturk et al. [28]
applied the DarkCovidNet deep learning model to raw chest X-ray
images and generated a binary classification (COVID-19, no find-
ings). They also used the DarkCovidNet classifier method on the
same dataset to create a triple classification (COVID-19, no-
findings, pneumonia). They declared that the highest accuracy of
the classifier was 98.08% for the binary classification and 87.02%
for the triple case. Sekeroglu et al. [29] proposed an alternative
COVID-19 detection method by using deep learning and machine
H. Arslan and H. Arslan Engineering Science and Technology, an International Journal 24 (2021) 839–847

learning classifiers on a publicly available dataset which contains k by introducing the generalized mean distance-based KNN classi-
1583 healthy, 4292 pneumonia and 225 confirmed COVID-19 chest fier. They stated that the proposed method is less sensitive to k
X-ray images. They stated that a convolutional neural network over the KNN-based classifier.
(CNN) without pre-processing and with minimized layers achieved
an accuracy of 98.50%. Jain et al. [30] developed a new approach to 3.2. Metrics
detect COVID-19 cases among chest X-ray images of healthy, bac-
terial pneumonia, viral pneumonia and COVID-19 by using To improve the performance of the KNN classifier, we investi-
ResNet18, ResNet101, DenceNet121 and VGG-16 deep learning gate the metrics used in the KNN algorithm under five categories
models. Their experimental results demonstrated that ResNet101 by following the similar categorization in [39,40]. We list them
method achieved the highest accuracy of 98.93%. Apostolopoulos as below:
and Bessiana [31] used VGG19, Mobile Net, Inception, Xception
and Inception ResNet v2 deep transfer learning classifiers to detect 1. L1 type metrics: Six metrics are investigated in this category.
COVID-19 positive cases among 224 COVID-19, 700 bacterial pneu- They are Manhattan metric (ManM), Chebyshev metric
monia and 504 normal chest X-ray images. Their results presented (ChebM), Canberra metric (CanM), Sorensen metric (SM),
that VGG19 achieved the best binary classification accuracy of Kulezynski metric (KM), and Mean character metric (MCM),
98.75% over the other CNN methods. which are explained in 1.
Ahuja et al. [32] used deep learning methods to detect COVID- 2. L2 type metrics: Seven metrics are investigated in this category.
19 positive cases from chest X-ray images. Asnaoui and Chawki They are Euclidean metric (EM), Clark metric (ClaM), Neyman
[33] constructed a novel method for detecting COVID-19 from v2 metric (NCSM), Squared v2 metric (SquM), Divergence metric
pneumonia chest X-ray and tomography images. They performed (DivM), and Squared Chi-squared metric (SCSM). We list these
recent deep learning methods, VGG16, VGG19, DenseNet201, metrics in 2.
Inception_ResNet_V2, Inception_V3, Resnet50, and MobileNet_V2. 3. Vicissitude metrics: Three metrics are considered in this cate-
They reported that the performance of Inception_Resnet_V2 gory. They are Vicis Symmetric 1 metric (VSDFM1), Vicis Sym-
demonstrated the best accuracy of 92.18%. Basu et al. [34] pro- metric 2 metric (VSDFM2) and Vicis Symmetric 3 metric
posed an alternative screening method of COVID-19, which is (VSDFM3). Definitions of the mentioned metrics are given in 3.
called Domain Extension Transfer Learning. They extracted some 4. Inner product metrics: Two metrics are investigated in this
discriminate features from the chest X-ray dataset. They then clas- category. They are Dice metric (DicM) and Chord metric
sified the images as normal, pneumonia, other diseases, and (ChoM). Definition of the mentioned metrics are demonstrated
COVID-19 with an accuracy of 95.3%. in 4.
5. Other metrics: In this category, we perform two metrics that
significantly affect the performance of the KNN algorithm.
3. Background
These metrics are Motyka (MotM) and Hassanat (HasM) met-
rics. The description of them is shown in 5, where
We propose a new COVID-19 detection method based on CpG ( 1þminðxi ;yi Þ
island features and the KNN classifier. The KNN is extremely useful 1  1þmaxðx i ;yi Þ
; if minðxi ; yi Þ P 0
Dðxi ; yi Þ ¼ 1þminðxi ;yi Þþjminðxi ;yi Þj .
for large data classification and its performance depends on the 1  1þmaxðx ;y Þþjminðx ;y Þj
; if minðxi ; yi Þ < 0
i i i i
distance metric used [35–38]. Several studies have been conducted
to detect optimum metrics for KNN algorithm in [39,40]. In this 4. Proposed method
section, first, we provide a brief description of the KNN method
and some recent studies based on the KNN. Second, we investigate In this section, we present a new COVID-19 detection method
the metrics used in the KNN classifier under five categories to based on the KNN classifier and CpG island features. Main steps
improve the performance of the KNN classifier on our model. of the proposed method are described in 1. The first step of the
algorithm is feature extraction. Extracting robust and discrimina-
3.1. K-nearest neighbors tive features from human coronaviruses genome sequences is the
most critical step to improve diagnosis of SARS-CoV-2. In this step,
K-nearest neighbors (KNN), a supervised machine learning algo- we propose to use CpG based features since CpG dinucleotides in
rithm, can be efficiently used to solve classification problems. The the open reading frames of SARS-CoV-2 has extremely low-
KNN was introduced in 1951 by [41] and then recasted in 1967 by frequency [46]. The main reason of lower CpG dinucleotides den-
[42]. The KNN is a non-parametric classifier known as one of the sity is the mutation of C into A and G into T [47,48]. The proposed
simplest and laziest algorithms. That is, there is no need to create features are extracted by using Eq. 1 and Eq. 2.
a learning model in this classification method. Despite the lazy CGp ¼ ratioðCÞ þ ratioðGÞ ð1Þ
structure of KNN, it was proposed as one of the 10 most effective
methods in the process of analyzing information in a database
given [43]. In the prediction process, the class to which a new
observation data belongs is determined by calculating the shortest Table 1
distance between the observation sample and its K-nearest neigh- L1 type metrics.
bors samples.
Abbreviation Metric Definition
There are some recent studies to improve the performance of Pn
i¼1 jxi  yi j
ManM Manhattan metric
the KNN. To decrease the sensitivity of the neighborhood size of
ChebM Chebyshev metric max jxi  yi j
k and improve voting strategy in the region of neighborhoods, 16i6n
CanM Canberra metric Pn jxi yi j
Gou et al. [44] proposed two k-nearest neighbor rules, which are i¼1 jx jþjy j
Pn i i

the weighted representation-based k-nearest neighbor rule and SM Sorensen metric jx y j

Pni¼1 i i
ðx þy Þ
the weighted local mean representation-based k-nearest neighbor Pi¼1
n
i i

KM Kulezynski metric jx y j
rule. Their experimental results demonstrate that the proposed Pn i¼1 i i
minðxi ;yi Þ
methods have a lower sensitiveness to k. Gou et al. [45] proposed Pn
i¼1

MCM Mean character metric i¼1

jxi yi j
another study to improve the selection of the neighborhood size of n

841
H. Arslan and H. Arslan Engineering Science and Technology, an International Journal 24 (2021) 839–847

Table 2 Algorithm 1. Proposed Algorithm

L2 type metrics.

Abbreviation Metric Definition Require:

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P
EM Euclidean metric S: training genome sequences,
ð ni¼1 jxi  yi j2 Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn L: COVID-19 positive, COVID-19 negative
ClaM Clark metric xi yi 2
i¼1 ðjxi jþjyi jÞ seq: a test sequence, and k: the neighborhood size
NCSM Neyman v2 metric Pn ðxi yi Þ2
i¼1 xi Ensure:
SquM Squared v2 metric Pn ðxi yi Þ2 Determine the class label of the test sequence seq
i¼1 xi þyi
DivM Divergence metric P 2 Step 1: Feature Extraction
2 ni¼1 ðxi yi Þ2
ðxi þyi Þ
Pn ðxi yi Þ2 1: for all genome sequences do
SCSM Squared Chi-squared metric
i¼1 jxi þyi j 2: feature1 ¼ ratioðCÞ þ ratioðGÞ
3: feature2 ¼ ratioðCGÞ=ratioðCÞratioðGÞ
4: end for
Table 3 Step 2: Apply KNN and use L1 type metric
Vicissitude metrics. 5: Compute the distance between seq and
Abbreviation Metric Definition every sample in S using any L1 type metric
VSDFM1 Vicis Symmetric 1 metric Pn ðxi yi Þ2
6: Choose k samples in S that are nearest to seq
i¼1 minðx ;y Þ2
i i 7: Assign seq to majority class
VSDFM2 Vicis Symmetric 2 metric Pn ðxi yi Þ2
i¼1 minðxi ;yi Þ
VSDFM3 Vicis Symmetric 3 metric Pn ðxi yi Þ2
i¼1 maxðxi ;yi Þ

5. Experimental results
Table 4
Inner product metrics. In this section, we evaluate performance of the proposed
Abbreviation Metric Definition method. Before discussing the results, first we explain our dataset,
Pn second we mention experimental setup, and then we provide a
DicM Dice metric 2 xi y
1  Pn 2 i¼1Pni 2 brief information about the performance measures. Finally, we will
x þ y
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
i¼1 i
P
i¼1 i

ChoM Chord metric n

xy
give the results of the experimental observations conducted by
2  2 Pn P
i¼1 i i
x2
i¼1 i
n
y2
i¼1 i
using different distance metrics.

Table 5
Other metrics: Motyka and Hassanat metrics.
Abbreviation Metric
MotM Motyka metric
HasM Hassanat metric
ratioðCGÞ
CpGo ¼
ratioðCÞratioðGÞ
where ratioðCÞ; ratioðGÞ, and ratioðCGÞ are computed as divided the
number of occurrences of C; G, and CG in the sequence, respectively
by the sequence length. Thus, each sequence containing 30 000
nucleotides is represented by two features only. 1 provides an
example of how features are calculated from a part of the sequence.
After feature extraction step, we apply the KNN method to clas-
sify SARS-CoV-2 sequences. The performance of the KNN mainly
depends on the metric that is used to compute the distances
between different data samples. To improve the performance of
the KNN algorithm, we perform 19 distance metrics investigated
in five categories, which are L1 type, L2 type, vicissitude, inner pro-
duct metrics, and the other types of metrics. In this step, we pro-
pose to use L1 type metric as a distance measure in the KNN.
5.1. Dataset
The 2019 novel coronavirus resource (2019nCoVR) [49] at China
National Center for Bioinformation is one of the most important
Definition
Pn sources of various types of coronaviruses. It integrates various
maxðxi ;yi Þ
Pi¼1n important databases including the GISAID, NCBI, NMDC and
ðx þyi Þ
i¼1 i
Pn CNCB/NGDC. In this study, we used complete genomic sequences
i¼1 Dðxi ; yi Þ
of human coronaviruses obtained from 2019nCoVR. Genome
sequence of each coronavirus has approximately a length of 30
000 nucleotides. The properties of human coronavirus sequences
ð2Þ are presented in 6. The dataset includes various types of human
coronaviruses such as AlphaCoV, BetaCov-1, MERS-CoV, NL63-
CoV, HKU1-CoV and 229E-CoV as well as SARS-COV-2. We refer
SARS-CoV-2 sequences as COVID-19 positive, and the sequences
that do not include SARS-CoV-2 are referred as COVID-19 negative.
In addition to 1000 SARS-CoV-2 sequences, we used 592 genome
sequences of other human coronaviruses in our experiments. We
note that all available genome sequences of human coronaviruses,
which are different from SARS-CoV-2 are downloaded.
5.2. Experimental setup
The experiments were performed using a core i7, 2.7 GHz pro-
cessor, 16 GB RAM under Linux operating system. Feature extrac-
Table 6
The properties of complete genome sequences of human coronaviruses.

Human coronaviruses The number sequences Label

SARS-CoV-2 1000 Positive
AlphaCoV 88 Negative
BetaCoV-1 140 Negative
MERS-CoV 258 Negative
NL63-CoV 61 Negative
Fig. 1. CpG based features. The numbers of C, G, and CG are 13, 20, and 3,
HKU1-CoV 18 Negative
respectively. Thus, CGp = ratio(C) + ratio(G) = 0.55, and CpGo = ratio(CG)/(ratio(C)
229E-CoV 27 Negative
ratio(G)) = 0.69.

842
H. Arslan and H. Arslan
tion and the classification process were performed using Python
programming language. Precision, recall, F-measure, and accuracy
in predicting COVID-19 positive case are used as the major perfor-
mance measures. Next, we define these performance measures.
5.3. Performance measures
COVID-19 prediction as a binary classification problem has four
prediction outcomes. True Positive (TP) is the number of genomic
sequences which are correctly classified as COVID-19 positive, True
Negative (TN) is the number of genome sequences which are cor-
rectly classified as COVID-19 negative, False Positive (FP) is the
number of genome sequences which are incorrectly classified as
COVID-19 positive, and False Negative (FN) is the number of
sequences which are incorrectly classified as COVID-19 negative.
The performance evaluation metrics such as precision, recall, F-
measure, and accuracy are defined by using these outcomes and
presented in confusion matrix by assigning actual and predicted
labels in 2. Next, we briefly explain performance parameters used
in this study.
Precision is the accuracy of the classifier in the presence of false
positive case. It is computed as the ratio of the number of correctly
classified positive samples to the number of samples labeled by the
system as positive and it is exhibited in Eq. 3.
TP
Precision ¼ ð3Þ
TP þ FP
Recall refers to as the number of positive class predictions made
out all positive samples in the data set. It is computed as in Eq. 4.
TP
Recall ¼ ð4Þ
TP þ FN
F-Measure is determined by the harmonic mean of precision
and recall, and computed using Eq. 5.
Precision  Recall
F  Measure ¼ 2  ð5Þ
Precision þ Recall
Accuracy is computed by dividing the total number of true cases
by all cases and it is indicated in Eq. 6.
TP þ TN
Accuracy ¼ ð6Þ
TP þ TN þ FP þ FN
5.4. Evaluation of the proposed features and KNN classifier with
respect to metrics
In this section, we evaluate the efficiency of proposed features
and the performance of the KNN classifier with respect to metrics
using a fivefold cross-validation that was performed on the data-
set. The dataset is randomly divided into two sets which are train-
ing and testing. Eighty percent of the entire human genome
sequences is used for training and the remaining 20% of the dataset
Fig. 2. Confusion matrix.
843
Engineering Science and Technology, an International Journal 24 (2021) 839–847
Fig. 3. 5-fold cross validation.
is used for validation. The experiments are repeated five times as
presented in 3. In our experiments, we increase the neighborhood
size of k between 1 and 20, and observe that the classification per-
formance of the proposed method remains the same or slightly
decreases with increasing the neighborhood size of k. For this rea-
son, in all experiments, the k value is basically set to 1.
7 presents precision, recall, and F-measure values of the KNN
classifier with respect to five metric groups, which are L1 type, L2
type, vicissitude, inner product metrics, and the other types of met-
rics. The results of the metrics in the same group are close to each
other; thus, we present the average of the results of the experi-
ments performed by the metrics in the same group. First, we ana-
lyze the results of the KNN equipped with L1 type metrics, which
are KM, ChebM, ManM, SM, CanM, and MCM. The KNN with L1
type metrics presents the best result, and achieves a precision of
98.4%, a recall of 99.2%, and an F-measure of 98.8%. This means that
almost all sequences are classified correctly. Next we look at the
results of the KNN with L2 type metrics, which are ClaM, EM,
SquM, NCSM, DivM, and SCSM. The KNN classifier with this group
metrics achieves a precision of 96.0%, a recall of 98.2%, and an F-
measure of 97.1% on average. When we look at the results of the
KNN with vicissitude metrics, it presents better result than KNN
with L2 type metrics. It achieves a precision of 98.4%, a recall of
99.0%, and an F-measure of 98.7%. Next, we investigate the results
of the KNN with inner product metrics, ChoM and DicM. This group
of metrics presents the worst results among the other groups, and
achieves a precision of 94.4%, a recall of 92.8%, and an F-measure of
93.4%. Finally, we investigate the other type of metrics, which are
HasM and MotM. The results of KNN with HasM are close to the
results of the KNN with inner product type metrics, and achieves
a precision of 94.4%, a recall of 92.9%, and an F-measure of 93.4%.
On the other hand, the results of the KNN with MotM is close to
them of L2 type metrics, and are better than that of HasM. It
achieves a precision of 96.4%, a recall of 98.2%, and an F-measure
of 97.3%.
In addition to precision, recall, and F-measure values, we pre-
sent accuracy results. We take the average of the accuracies of
KNN classifier obtained with the metrics in the same group, and
show the results for each metric group separately in 4. The accu-
racy results of the KNN classifier using inner product metrics and
HasM are close to each other and have the worst accuracy, 91.7%,
and 91.8%, respectively. When L2 type metrics are used, the
method achieves an accuracy of 96.2%. The KNN with MotM metric
has remarkable accuracy values, which is 96.5%. The accuracy val-
ues of the KNN with L1 and Vicissitude type metrics are close to
each other, and our method achieves the best accuracy with
98.4% when L1 type metrics are used.
H. Arslan and H. Arslan Engineering Science and Technology, an International Journal 24 (2021) 839–847

Table 7
Average of the precision, recall, and F-measure values.

Group Metric name Precision (%) Recall (%) F-measure (%)

L1 KM 98.4 99.2 98.8
ChebM
ManM
SM
CanM
MCM

L2 ClaM 96.0 98.2 97.1

EM
SquM
NCSM
DivM
SCSM

Vicissitude VSDFM1 98.4 99.0 98.7

VSDFM2
VSDFM3

Inner Product ChoM 94.4 92.8 93.4

DicM

Other HasM 94.4 92.9 93.4

MotM 96.4 98.2 97.3

of the results in terms of the method, dataset, class, and accuracy.

Fig. 4. Accuracy of the proposed method with respect to each group of metric.
6. Comparisons with other methods
In this part, we compare the results of the proposed method
with the state-of-the-art methods. 8 provides a simple comparison
We investigate these studies under two categories: the studies
using genome sequences datasets and the other datasets.
First, we compare our method with the studies which detect
COVID-19 cases from genome sequences. Randhawa et al. [23]
combined supervised machine learning methods with digital signal
processing. They used 29 sequences of SARS-CoV-2, and 20
sequences for each of alphacoronavirus, betacoronavirus, and
deltacoronavirus. They achieved 100% accuracy when linear dis-
criminant method was used. Main limitation of their proposed
method is the number of sequences used in their dataset. Further-
more, Randhawa et al. used delta coronaviruses genomes. How-
ever, deltacoronavirus mainly causes an infectious disease among
bird species rather than human. When the number of sequences
in their dataset is increased and human coronavirus sequences
genetically similar to SARS-CoV-2 sequences are added to their
dataset, overall accuracy of their method may decrease. Another
study predicting COVID-19 from genome sequences is introduced
by Naeem et al. [24]. They used the classical KNN method to distin-
guish SARS-CoV-2 sequences among SARS-CoV-2, SARS-CoV and
MERS-CoV genome sequences. They used 76 sequences for each
of SARS-CoV-2, SARS-CoV, and MERS-CoV, and achieved an accu-
racy of 100%. They extracted features by using Discrete Fourier
transform, Discrete cosine transform, and Seven Moment Invari-
ants methods. When comparing to our method, their feature

844
H. Arslan and H. Arslan Engineering Science and Technology, an International Journal 24 (2021) 839–847

Table 8
Comparison of existing state-of-the-art classification studies.

Study Method used Dataset Classes Accuracy (%)

Randhawa et al.[23] Linear Discriminant 29 COVID-19 virus sequences COVID-19 100
Linear SVM 20 alphacoronavirus sequences vs other types of viruses
Quadratic SVM 20 betacoronavirus sequences
Fine KNN 20 deltacoronavirus sequences
Subspace Discriminant
Subspace KNN

Naeem et al. [24] Discrete Fourier Transform with KNN 76 COVID-19 sequences COVID-19 100
Discrete Cosine transform with KNN 76 SARS-CoV sequences vs SARS-CoV
Seven Moment Invariants with KNN 76 MERS-CoV sequences vs MERS-CoV

Proposed Method KNN with L1 type metrics 1000 COVID-19 sequences COVID-19 vs 98.4
CpG based features 592 other coronavirus sequences other type coronaviruses

Unal and Dudak [26] Naive Bayes 95839 case with COVID-19 vs 100
KNN 19 type of laboratory findings other diseases
Support Vector Machines
Decision Tree Algorithm

Barstugan et al. [27] Support Vector Machine 150 Abdominal CT images COVID-19 vs 99.6
other viral pneumonias

Apostolopoulos and Bessiana [31] Deep Transfer Learning 224 COVID-19 Chest X-ray images COVID-19 vs 98.75
700 images of bacterial pneumonia Pneumonia vs
504 images of normal conditions Normal

Ozturk et al. [28] DarkCovidNet raw chest X-ray images COVID-19 vs 87.02
healthy vs
pneumonia

Jain et al. [30] VGG16 chest X-ray images COVID-19 vs 98.93

ResNet18 bacterial pneumonia vs
DenceNet121 viral pneumonia vs
ResNet101 healthy

(continued on next page)

845
H. Arslan and H. Arslan
Table 8 (continued)
Study Method used
Asnaoui and Chawki [33] VGG16
VGG19
DenseNet201
Inception ResNet V2
Inception V3
Resnet50
MobileNet V2
Basu et al. [34] Domain Extension Transfer Learning
extraction method is expensive. Another disadvantage of the
model established by Naeem et al. [24] is that they work with a
smaller dataset just like Randhawa.
Next we discuss advantages/drawbacks of the methods using
the other datasets. Unal and Dudak [26] used Mexico Patient
Health Dataset to detect COVID-19 cases. Although the size of their
dataset is large, their dataset only represents the features of a
specific region. Furthermore, they used 19 different features
including the sex of the patient, age of the patient, the state of
pneumonia and intubation as well as the state of many other dis-
eases. One of the advantages of our model over the method of Unal
and Dudak [26] is to use powerful and effective two features
derived from the complete genome sequences of human coron-
aviruses. Thus, our proposed method detects COVID-19 positive
cases within a few seconds. However, Unal and Dudak [26] did
not provide any explanation about the detection time. Another dif-
ferent aspects of the proposed method from their method is that
they use the classicial KNN classifier with default parameters
although we use the KNN classifier with an optimum distance met-
ric in our proposed model.
There are a number of the existing studies detecting COVID-19
cases from image dataset in the literature. In 8, we also exhibit six
image based studies achieving remarkable accuracy values. These
image based methods used different deep learning or machine
learning methods. Although image based studies have a higher
accuracy, when considering a high mutation rate of SARS-CoV-2,
using genomic sequences is extremely beneficial when tracking
coronavirus genes that change frequently as the disease spreads
from one person to another. Moreover, the radiation generated
by X-ray or CT scanning machines may cause permanent damages
to people. For this reason, X-ray or CT scanning may not be
obtained for some people, which can be considered as a disadvan-
tage of image based studies.
7. Conclusion
COVID-19 is an existing epidemic that sets new records in terms
of cumulative and daily numbers for global infection. The pan-
demic is an unprecedented situation in healthcare systems world-
wide, and to overcome this pandemic, it is essential to accurately
detect COVID-19 cases by analyzing data of patients in a minimum
amount of time. In this study, we propose an accurate and fast
method to detect COVID-19 positive cases from genome sequences
of human coronaviruses. In the proposed method, first, the features
that significantly differentiate COVID-19 cases are extracted from
the complete genome sequences of human coronaviruses. In this
Engineering Science and Technology, an International Journal 24 (2021) 839–847
Dataset Classes Accuracy (%)
1583 normal X-ray and CT images COVID-19 vs 92.18
231 COVID-19 images other viral pneumonias
2780 images of bacterial pneumonia
1493 coronavirus images
chest X-ray images COVID-19 vs 95.3
normal vs
pneumonia vs
other diseases
step, we propose to use CpG island features. Each genome
sequence of human coronavirus, which includes about 30,000
nucleotides is represented by two real numbers only. Feature
extraction step takes just a few seconds. Second, the KNN method
is used for the classification of COVID-19 positive cases from the
other types of human coronaviruses, AlphaCoV, Beta-Cov-1,
MERS-CoV, NL63-CoV, HKU1-CoV, and 229E-CoV. The KNN classi-
fier is the simplest method and has high flexibility for solving com-
plex classification problems. The accuracy of the KNN is higher
than state-of the-art classifiers in certain cases, and it often pro-
duces efficient performance. However, the performance of the
KNN greatly depends on the metric performed. To detect the most
appropriate metric, we review five groups of metrics used in the
KNN classifier. The selection of different distance metrics for the
KNN can result a variation in accuracy outcomes for the same data-
set. Experimental results reveal that the proposed method achieves
the highest accuracy, which is 98.4% on average in a few seconds
when L1 type metrics are used as a distance measure in the KNN.
In future studies, we will compare human SARS-CoV-2 sequences
to other types of coronavirus sequences such as bat SARS-CoV-
like coronaviruses 2, and propose a similarity based feature to
increase overall accuracy. In addition, the factors affecting the
recovery status of patients suffering from COVID-19 may be inves-
tigated in future studies by combining machine learning and paral-
lel computing methods with effective features.
Declaration of Competing Interest
The authors declare that they have no known competing finan-
cial interests or personal relationships that could have appeared
to influence the work reported in this paper.
References
[1] D. Schoeman, B. Fielding, Coronavirus envelope protein: current knowledge,
Virol. J. 16. doi: 10.1186/s12985-019-1182-0..
[2] R.J. de Groot, S. Baker, R. Baric, L. Enjuanes, A. Gorbalenya, K. Holmes, S.
Perlman, L. Poon, P. Rottier, P. Talbot, et al., Family coronaviridae, Virus Taxon.
(2012) 806–828.
[3] J. Cui, F. Li, Z.L. Shi, Origin and evolution of pathogenic coronaviruses, Nat. Rev.
Microbiol. 17 (2019) 181–192, https://doi.org/10.1371/journal.pone.0119815.
[4] A chronicle on the sars epidemic, Chin. Law Govern. 36 (4) (2003) 12–15.
arXiv:https://doi.org/10.2753/CLG0009-4609360412, doi: 10.2753/CLG0009-
4609360412..
[5] A. Zumla, D.S. Hui, S. Perlman, Middle east respiratory syndrome, The Lancet
386 (9997) (2015) 995–1007, https://doi.org/10.1016/S0140-6736(15)60454-
8. http://www.sciencedirect.com/science/article/pii/S0140673615604548.
[6] L. Xingguang, W. Wei, Z. Xiaofang, Z. Junjie, Z. Qiang, L. Yi, C. Antoine,
Transmission dynamics and evolutionary history of 2019-ncov, J. Med. Virol.
92 (2020) 501–511, https://doi.org/10.1002/jmv.25701.
846
H. Arslan and H. Arslan
[7] L. Fu, B. Wang, T. Yuan, X. Chen, Y. Ao, T. Fitzpatrick, P. Li, Y. Zhou, Y. fan Lin, Q.
Duan, G. Luo, S. Fan, Y. Lu, A. Feng, Y. Zhan, B. Liang, W. Cai, L. Zhang, X. Du, L. Li,
Y. Shu, H. Zou, Clinical characteristics of coronavirus disease 2019 (covid-19) in
china: A systematic review and meta-analysis, J. Infect. 80 (6) (2020) 656–665.
doi: https://doi.org/10.1016/j.jinf.2020.03.041. URL http://
www.sciencedirect.com/science/article/pii/S0163445320301705..
[8] R. Vaishya, M. Javaid, I. Khan, A. Haleem, Artificial intelligence (ai) applications
for covid-19 pandemic, Diab. Metab. Syndrome Clin. Res. Rev. 14. doi: 10.1016/
j.dsx.2020.04.012..
[9] G.G. Waleed Salehi A, Baglat P, Review on machine and deep learning models
for the detection and prediction of coronavirus, Mater. Today Proc. doi:
10.1016/j.matpr.2020.06.245..
[10] M.A. Dey L., S. Chakraborty, Machine learning techniques for sequence-based
prediction of viral–host interactions between sars-cov-2 and human proteins,
Biomed. J. doi: 10.1016/j.bj.2020.08.003. URL https://doi.org/10.1016/j.
bj.2020.08.003..
[11] P. Gupta, A. Sharma, R. Jindal, Scalable machine-learning algorithms for big
data analytics: a comprehensive review, WIREs Data Min. Knowl. Discovery 6
(6) (2016) 194–214. arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1002/
widm.1194, doi: 10.1002/widm.1194. https://onlinelibrary.wiley.com/doi/
abs/10.1002/widm.1194..
[12] M.M. Najafabadi, F. Villanustre, T.M. Khoshgoftaar, N. Seliya, R. Wald, E.
Muharemagic, Deep learning applications and challenges in big data analytics,
J. Big Data 2 (1) (2015) 194–214, https://doi.org/10.1186/s40537-014-0007-7.
[13] T.T. Zin, J.C.W. Lin, Big data analysis and deep learning applications:
proceedings of the first international conference on big data analysis and
deep learning, 2019. URL https://www.springer.com/gp/book/
9789811308680.
[14] T. Das, Machine learning algorithms for image classification of hand digits and
face recognition dataset, Int. Res. J. Eng. Technol. 4 (12) (2017) 640–649.
[15] M. Sharma, J. Anuradha, H. Manne, G. Kashyap, Facial detection using deep
learning, IOP Conf. Ser.: Mater. Sci. Eng. 263 (2017), https://doi.org/10.1088/
1757-899X/263/4/042092 042092.
[16] P. Viola, M. Jones, Robust real-time face detection, Int. J. Comput. Vision 57
(2004) 137–154, https://doi.org/10.1023/B:VISI.0000013087.49260.fb.
[17] K. Lewandowski, Y. Xu, S. Pullan, S. Lumley, D. Foster, N. Sanderson, A.
Vaughan, M. Morgan, N. Bright, J. Kavanagh, R. Vipond, M. Carroll, A. Marriott,
K. Gooch, M. Andersson, K. Jeffery, T. Peto, D. Crook, A. Walker, P. Matthews,
Metagenomic nanopore sequencing of influenza virus direct from clinical
respiratory samples, J. Clin. Microbiol. 58. doi: 10.1128/JCM.00963-19..
[18] L. Kafetzopoulou, K. Efthymiadis, K. Lewandowski, A. Crook, D. Carter, J.
Osborne, E. Aarons, R. Hewson, J. Hiscox, M. Carroll, R. Vipond, S. Pullan,
Assessment of metagenomic nanopore and illumina sequencing for recovering
whole genome sequences of chikungunya and dengue viruses directly from
clinical samples, Eurosurveillance 23. doi: 10.2807/1560-7917.
ES.2018.23.50.1800228..
[19] A. Khanday, S. Rabani, Q. Khan, et al., Machine learning based approaches for
detecting covid-19 using clinical text data, Int. J. Inf. Technol. 12 (2020) 731–
739, https://doi.org/10.1007/s41870-020-00495-9.
[20] L. Li, L. Qin, Z. Xu, Y. Yin, X. Wang, B. Kong, J. Bai, Y. Lu, Z. Fang, Q. Song, K. Cao,
D. Liu, G. Wang, Q. Xu, X. Fang, S. Zhang, J. Xia, J. Xia, Using artificial
intelligence to detect covid-19 and community-acquired pneumonia based on
pulmonary ct: evaluation of the diagnostic accuracy, Radiology 296 (2) (2020
Aug) E65–E71, pMID: 32191588. doi: 10.1148/radiol.2020200905. doi:
10.1007/s41870-020-00495-9..
[21] A. Alimadadi, S. Aryal, I. Manandhar, P.B. Munroe, B. Joe, X. Cheng, Artificial
intelligence and machine learning to fight covid-19, Physiol. Genom. 52 (4) (2020)
200–202, pMID: 32216577. arXiv:https://doi.org/10.1152/physiolgenomics.
00029.2020, doi: 10.1152/physiolgenomics.00029.2020. doi: 10.1152/
physiolgenomics.00029.2020..
[22] J. Chen, K. Li, Z. Zhang, K. Li, P.S. Yu, A Survey on Applications of Artificial
Intelligence in Fighting Against COVID-19, arXiv e-prints (2020)
arXiv:2007.02202arXiv:2007.02202..
[23] G.S. Randhawa, M.P.M. Soltysiak, H. El Roz, C.P.E. de Souza, K.A. Hill, L. Kari,
Machine learning using intrinsic genomic signatures for rapid classification of
novel pathogens: Covid-19 case study, vol. 15, Public Library of Science, 2020,
pp. 1–24. doi: 10.1371/journal.pone.0232391. URL https://doi.org/10.1371/
journal.pone.0232391..
[24] S.M. Naeem, M.S. Mabrouk, S.Y. Marzouk, M.A. Eldosoky, A diagnostic genomic
signal processing (GSP)-based system for automatic feature analysis and
detection of COVID-19, Brief. Bioinf. Bbaa170. arXiv:https://academic.
oup.com/bib/advance-article-pdf/doi/10.1093/bib/bbaa170/33650674/
bbaa170.pdf, doi: 10.1093/bib/bbaa170. URL https://doi.org/10.1093/bib/
bbaa170..
[25] A.F.d.M. Batista, J.L. Miraglia, T.H.R. Donato, A.D.P. Chiavegatto Filho, Covid-19
diagnosis prediction in emergency care patients: a machine learning approach,
847
Engineering Science and Technology, an International Journal 24 (2021) 839–847
medRxivdoi: 10.1101/2020.04.04.20052092. https://www.medrxiv.
org/content/early/2020/04/14/2020.04.04.20052092..
[26] Y. Ünal, M.N. Dudak, Classification of covid-19 dataset with some machine
learning methods (2020)..
[27] M. Barstugan, U. Ozkaya, S. Ozturk, Coronavirus (covid-19) classification using
ct images by machine learning methods. arxiv:2003.09424 (03 2020)..
[28] T. Ozturk, M. Talo, E.A. Yildirim, U.B. Baloglu, O. Yildirim, U. Rajendra Acharya,
Automated detection of covid-19 cases using deep neural networks with X-ray
images, Comput. Biol. Med. 121 (2020), https://doi.org/10.1016/
j.compbiomed.2020.103792 103792.
[29] B. Sekeroglu, I. Ozsahin, Detection of covid-19 from chest x-ray images using
convolutional neural networks, SLAS TECHNOL. Transl. Life Sci. Innov. 0 (0) (0)
2472630320958376, pMID: 32948098. arXiv:https://doi.org/10.1177/
2472630320958376, doi: 10.1177/2472630320958376. doi: 10.1177/
2472630320958376..
[30] G. Jain, D. Mittal, D. Thakur, M.K. Mittal, A deep learning approach to detect
covid-19 coronavirus with X-ray images, Biocybern. Biomed. Eng. 40 (4)
(2020) 1391–1405, https://doi.org/10.1016/j.bbe.2020.08.008.
[31] I. Apostolopoulos, T. Bessiana, Covid-19: automatic detection from X-ray
images utilizing transfer learning with convolutional neural networks, Phys.
Eng. Sci. Med. 43 (2) (2020) 635–640, https://doi.org/10.1007/s13246-020-
00865-4.
[32] S. Ahuja et al., Deep transfer learning-based automated detection of covid-19
from lung ct scan slices, Appl. Intell. (2020), https://doi.org/10.1007/s10489-
020-01826-w.
[33] K.E. Asnaoui, Y. Chawki, Using X-ray images and deep learning for automated
detection of coronavirus disease, J. Biomol. Struct. Dyn. 0 (0) (2020) 1–12,
pMID: 32397844. arXiv:https://doi.org/10.1080/07391102.2020.1767212, doi:
10.1080/07391102.2020.1767212. doi: 10.1080/07391102.2020.1767212..
[34] S. Basu, S. Mitra, N. Saha, Deep learning for screening covid-19 using chest X-
ray images (2020). arXiv:2004.10507..
[35] J. Maillo, I. Triguero, F. Herrera, A mapreduce-based k-nearest neighbor
approach for big data classification, in: 2015 IEEE Trustcom/BigDataSE/ISPA,
vol. 2, 2015, pp. 167–172..
[36] Z. Deng, X. Zhu, D. Cheng, M. Zong, S. Zhang, Efficient knn classification
algorithm for big data, Neurocomput. 195 (C) (2016) 143–148, https://doi.org/
10.1016/j.neucom.2015.08.112.
[37] J. Maillo, S. Ramírez, I. Triguero, F. Herrera, knn-is: an iterative spark-based
design of the k-nearest neighbors classifier for big data, Knowl.-Based Syst.
117. doi: 10.1016/j.knosys.2016.06.012..
[38] F. Wang, Q. Wang, F. Nie, W. Yu, R. Wang, Efficient tree classifiers for large
scale datasets, Neurocomputing doi: 10.1016/j.neucom.2017.12.061..
[39] S.H. Cha, Comprehensive survey on distance/similarity measures between
probability density functions, Int. J. Math. Models Methods Appl. Sci. 1 (4) (20)
300–307..
[40] H. Abu Alfeilat, A. Hassanat, O. Lasassmeh, A. Tarawneh, M. Alhasanat, H. Eyal-
Salman, S. Prasath, Effects of distance measure choice on k-nearest neighbor
classifier performance: a review, Big Data 7. doi: 10.1089/big.2018.0175..
[41] E. Fix, J.L.H. (1951), Discriminatory analysis. nonparametric discrimination:
consistency properties, Technical Report 4, USAF School of Aviation Medicine,
Randolph Field, TX, USA. URL http://www.jstor.org/stable/1403797..
[42] T. Cover, P. Hart, Nearest neighbor pattern classification, IEEE Trans. Inf.
Theory 13 (1)..
[43] X. Wu, V. Kumar, J. Quinlan, J. Ghosh, Q. Yang, H. Motoda, G.J. McLachlan, A.F.
M. Ng, B. Liu, P. Yu, Z. Zhou, M. Steinbach, D. Hand, D. Steinberg, Top 10
algorithms in data mining, Knowl. Inf. Syst. 14 (2007) 1–37.
[44] J. Gou, W. Qiu, Z. Yi, X. Shen, Y. Zhan, W. Ou, Locality constrained
representation-based k-nearest neighbor classification, Knowl.-Based Syst.
167 (2019) 38–52, https://doi.org/10.1016/j.knosys.2019.01.016. http://
www.sciencedirect.com/science/article/pii/S0950705119300152.
[45] J. Gou, H. Ma, W. Ou, S. Zeng, Y. Rao, H. Yang, A generalized mean distance-
based k-nearest neighbor classifier, Expert Syst. Appl. 115 (2019) 356–372,
https://doi.org/10.1016/j.eswa.2018.08.021. http://
www.sciencedirect.com/science/article/pii/S0957417418305293.
[46] X. Xia, Extreme genomic CpG deficiency in SARS-CoV-2 and evasion of host
antiviral defense, Mol. Biol. Evol. 37 (9) (2020) 2699–2705. arXiv:https://
academic.oup.com/mbe/article-pdf/37/9/2699/33721082/msaa094.pdf, doi:
10.1093/molbev/msaa094. doi: 10.1093/molbev/msaa094..
[47] Y. Wang, J.M. Mao, G.D. Wang, Z.P. Luo, L. Yang, Q. Yao, K.P. Chen, Human sars-
cov-2 has evolved to reduce cg dinucleotide in its open reading frames, Sci.
Rep. 10 (2020) 5165–5184.
[48] H. Dinka, A. Milkesa, Unfolding sars-cov-2 viral genome to understand its gene
expression regulation, Infect. Genet. Evol. 84. doi: 10.1016/j.
meegid.2020.104386..
[49] The 2019 novel coronavirus resource, https://bigd.big.ac.cn/ncov, accessed:
2020-09-24..