The International Journal of Multimedia & Its Applications (IJMA) Vol.14, No.

5, October 2022

GPCR PROTEIN FEATURE REPRESENTATION USING

DISCRETE WAVELET TRANSFORM AND PARTICLE
SWARM OPTIMISATION ALGORITHM
Nor Ashikin Mohamad Kamal1, Azuraliza Abu Bakar2 and Suhaila Zainudin2
1
School of Computing Sciences, College of Computing, Informatics and Media,
Universiti Teknologi MARA, 40450 Shah Alam, Selangor, Malaysia.
2
Centre for Artificial Intelligence Technology (CAIT), Faculty of Information
Science and Technology, Universiti Kebangsaan Malaysia, 43600 Bangi,
Selangor, Malaysia.

ABSTRACT
Features play an important role in representing classes in the hierarchy structure, and using unsuitable
features will affect classification performance. The discrete wavelet transform (DWT) approach provides the
ability to create the appropriate features to represent data. DWT can produce global and local features using
different wavelet families and decomposition levels. These two parameters are essential to obtain a suitable
representation for classes in the hierarchy structure. This study proposes using a particle swarm optimisation
(PSO) algorithm to select the suitable wavelet family and decomposition level for G-protein coupled receptor
(GPCR) hierarchical class representation. The results indicate that the PSO algorithm mostly selects
Biorthogonal wavelets and decomposition level 2 to represent GPCR protein. Concerning the performance,
the proposed method achieved an accuracy of 97.9%, 85.9%, and 77.5% at the family, subfamily, and sub-
subfamily levels, respectively.

KEYWORDS
Decomposition level, GPCR, hierarchical classification, particle swarm optimisation, wavele tfamily

1. INTRODUCTION
Proteins play an important role in the structure and function of all living cells and viruses. Proteins
have several general functions: structural, defence, receptor, signal, transport, movement, hormone,
and storage. Receptors are protein molecules receiving chemical signals from outside the cell and
allowing certain molecules to enter and leave the cell. G-protein coupled receptors (GPCR) exist
on the surface of every cell [1]. GPCRs generate signals in cells to regulate key physiological
processes, such as hormone signalling, neurotransmission, cognition, vision, taste, pain perception,
and others. It is also known as the seven transmembrane domain (7TM) receptor because there are
seven transmembrane segments in which three loops are outside the cell, and three loops are inside
the cell, with the N-terminal position being inside the cell and the C- terminal being outside the
cell. GPCR consists of three levels: family, subfamily, and sub- subfamily. The family level consists
of five classes: families A, B, C, D, and E. The subfamily level consists of 38 classes, and the sub-
subfamily level consists of 87 classes. Since GPCR has very complicated relationships between
classes, classifying GPCR has proven to be very difficult [2]. In addition, many protein sequences
in the same family share homology with protein sequences in other families, increasing
classification difficulty [3]. GPCR classification not only depends on sequence order but also
includes structural, functional, and evolutionary

DOI:10.5121/ijma.2022.14501 1
 The International Journal of Multimedia & Its Applications (IJMA) Vol.14, No. 5, October 2022
characteristics, such as chemical and pharmacological factors [4]. GPCRs are one of the most
challenging data sets to classify based on these factors.

The discrete wavelet transform (DWT) method is suitable for representing features of biological
data ([5]; [6]; [7]; [8]; [9]). With the nature of the multire solution analysis, DWT can provide
information on protein sequences more effectively and allow biological signals to be analysed in
the frequency domain and time domain ([9]; [6]). This property is not found in signal processing
methods, such as the Fourier transform, which can only study signals in the frequency domain ([10];
[11]; [8]; [12]). Therefore, the advantages of this DWT method can provide more information than
other feature representation methods ([13]; [14]; [15]; [16]; [17]). DWT can produce global and
local features in various decomposition levels to be analysed and produce features that do not have
overlaps [18]. DWT can also decompose global features, such as PseAAC and AAC, into
coefficients at different decomposition levels, producing global and local features for a protein
sequence ([9]). Global features are obtained from approximation coefficients, while local features
are obtained from detailed DWT coefficients [6]. [19] used DWT Coiflet 4 family with
decomposition level 3 to obtain global and local features. Although this method has achieved high
accuracy for all three levels of the GPCR protein hierarchy, the dataset only includes classification
at the family, subfamily A, and sub-subfamily Amine levels.

Selecting the appropriate DWT family type and decomposition level is important in data analysis.
It is because an accurate representation will preserve the important data features to be analysed and
further help to understand its organisation and complexity ([20]; [21]). The selected DWT family
needs to meet the characteristics of orthogonality, symmetry, and shape similarity with the studied
data signal ([22]; [23], [25]). Nevertheless, most studies chose the family type and DWT
decomposition level selection based on experience or manually ([20]; [24]). Different family types
and decomposition levels can be used in DWT. Types of DWT families are Haar, Daubechies,
Coiflets, Symlets, Discrete Meyer, and Biorthogonal. The decomposition level is related to the
number of global and local features produced from DWT [26]. Using a high decomposition level
on a short sequence will produce overlapping information, while using a low decomposition level
on a long sequence will ignore much detailed information about the sequence ([14]; [27]).
Therefore, selecting the appropriate family type and DWT decomposition level for the feature
representation of a class is important because these two parameters affect the classification
performance ([28]; [29];[30]; [26]; [31]). Some studies used metaheuristic methods to optimise the
selection of family types and DWT decomposition levels, such as genetic algorithms ([32]; [33]),
particle swarm optimisation ([34]; [26], [32]), whale optimisation algorithm [35], and evolutionary
quantum swarm algorithm [36]. Many researchers used metaheuristic methods to obtain the optimal
wavelet family and the decomposition level for research in engineering fields. However, no
research has been done to choose the optimal DWT family type and decomposition level for protein
feature representation in hierarchical classes.

The organisation of the remaining paper is as follows. The methodology section discusses the
proposed methods, the particle swarm optimisation (PSO) algorithm to select the wavelet family
and decomposition level to represent each parent class in the G-protein coupled receptors (GPCR)
protein hierarchy. This section also explains the hierarchical classification and classifier used in
this study. Lastly, the results and discussions section contains the experiment’s findings, and
conclusions are presented in the last section.

2
 The International Journal of Multimedia & Its Applications (IJMA) Vol.14, No. 5, October 2022

2. METHODOLOGY
Figure 1 shows the flowchart of the overall GPCR protein hierarchical classification using the
features generated from the optimal wavelet family and decomposition level. This process begins
by representing the GPCR protein sequence using the pseudo amino acid composition algorithm
(PseAAC). This algorithm converts each protein sequence consisting of letters into a vector with
170 numerical features. Next, these numerical features must be converted to wavelet coefficients
using the suitable wavelet family, m, and decomposition level, n. There are 82 wavelet families,
and seven decomposition levels can be chosen. The PSO algorithm has been used to determine each
class’s wavelet family and decomposition level at the parent node of the GPCR protein hierarchy.

Figure 1. The GPCR protein hierarchical classification flowchart

The DWT process at each decomposition level, j, will produce approximation and detail
coefficients; they will be used for the transformation process for the next level. The number of
coefficients generated depends on the decomposition level, j, chosen by the PSO algorithm. This

3
 The International Journal of Multimedia & Its Applications (IJMA) Vol.14, No. 5, October 2022
feature vector is then classified using a top-down hierarchical classification algorithm using the
SVM classifier. The following subsections describe each step involved in the GPCR protein
hierarchical classification, as shown in Figure 1.

2.1. Data Collection

This study uses the G-Protein Coupled Receptor (GPCR) protein hierarchy benchmark dataset
collected by [37], known as GDS. The entire protein sequence has a length measurement of no
more than 280 residues. Only classes with a number of data greater than ten at the third hierarchical
level are used. It makes a total of five classes at the first level, 36 at the second level, and 87 at the
third level of the hierarchy. The total number of protein sequences used in this study is 8,222 protein
sequences. The GDS dataset is available at http://www.cs.kent.ac.uk/projects/biasprofs/down-
208loads.html. This data has been used by past researchers, such as [38], [39], [40], [41], [42], and
[43].

2.2. Feature Representation

This study uses two feature representation methods: pseudo-amino acid composition (PseAAC)
and discrete wavelet transform (DWT).

2.2.1. Pseudo Amino Acid Composition (PseAAC)

The GPCR protein sequences are first represented using the pseudo amino acid composition
(PseAAC) method. This method transforms the protein sequences into numerical features. Here are
the steps implemented to transform protein sequences into numerical features using the PseAAC
method. In the PseAAC method, protein P can be written as:

PseAAC=P1,P2,...,P20,...,P^ (1)
Where

The first twenty elements from P1 to P20 in Equation (1) represent the frequency of the amino acid
in the sequence. The symbol λ is the amino acid correlation rank used where λ= 1,…, m. This rank
value is a non-negative integer value smaller than the protein sequence’s length. The n value
represents the number of physiochemical properties of the amino acid used. This study uses six
physiochemical properties of the amino acid as in [44], namely the mass properties, the pK group
from pK from the α-COOH group, and the pI group at 25∘C. The weighting factor, w, is created to
emphasise PseAAC compared to protein features using the amino acid composition. PseAAC
protein features are obtained directly through the PseAAC web server ([45]). There are two types
of PseAAC: type I and type II. The study of [45] found that PseAAC type II is more suitable to be
used as a protein feature. PseAAC type II considers the contribution of physiochemical properties
during the calculation ([46]). Therefore, the features generated for each protein sequence from the
parameters determined above according to Equation (2) is = 20+6*25, which is 170 features.

The feature normalisation process is next on the PseAAC features, in which each feature is scaled
to the limit range [0,1] so that the feature value is in a small range. Normalising features can prevent
features with large values from dominating features with small values. In addition, it can also
simplify the calculation process. The formula for feature normalisation is in Equation (3):

(3)

4
 The International Journal of Multimedia & Its Applications (IJMA) Vol.14, No. 5, October 2022

2.2.2. Discrete Wavelet Transform (DWT)

The next step is calculating the wavelet transform on the normalised PseAAC features. There are
two important functions in DWT: the scaling and wavelet functions. The scaling function is used
to get low-frequency features for the normalised PseAAC features, and these low-frequency
features are known as the approximation coefficients containing global information. The definition
of the scaling functionis as in Equation (4) [47]:

(4)

where x, a, and b denote the data, scaling, and translation parameters, respectively. The wavelet
function is used to obtain detail coefficients or local features. The definition of the wavelet
function is in Equation (5) [47]:

(5)

This study combines approximation and detail coefficients depending on the wavelet family
selection and decomposition level selection to form the coefficient vector. The detail coefficients
are retained since they may contain valuable information. DWT can be implemented using different
wavelet family types. Each wavelet family type has its scaling and wavelet functions [31]. The right
family type should be selected so DWT can be used efficiently [18]. In this study, 82 wavelet
families are available to represent GPCR proteins. These wavelet families include 20 from the
Daubechies family (Db1-Db20), 20 from the Symlet family (Sym1-Sym20), five from the Coiflet
family (Coif1-Coif5), six from the Fejer-Korovk in family, 15 from the Biorthogonal family, 15
from the Reverse Biorthogonal family, and one from the Meyer family. Each family differs based
on compact support, vanishing moment, symmetry, and orthogonality [48].

DWT can analyse the PseAAC feature vectors at various resolutions, which can be done by
decomposing the PseAAC feature vector into several decomposition levels. The maximum level
that can be implemented is based on the formula log2 (N), where N is the size of the PseAAC
feature. Since the feature size of PseAAC is 170, the maximum possible decomposition is seven
levels.

2.2.3. Particle Swarm Optimisation (PSO) Algorithm

Particle swarm optimisation (PSO) is an intelligence-oriented, stochastic computing technique

introduced by Kennedy and Eberhart in 1995 [49]. PSO has been widely used for many optimisation
problems because of its unique search mechanism and simple implementation. It was inspired by
the social behaviour of birds looking for food. In this algorithm, the birds in the flock are a solution
in a high-dimensional space with four vectors: the current position, the best position obtained so
far, the best position of the other particles, and the velocity. According to [50, 51], the PSO
algorithm does not require many calculations and fast convergence. Two important parameters in
PSO are pbest and gbest. The pbest value is the best accuracy value for each particle up to iteration,
t, while the gbest value is the best global accuracy value. The fitness function for each particle is
calculated using Equation (6) where c(i) is the number of protein sequences correctly classified in
class i and Total (i) is the total number of protein sequences in class i. Each particle updates its
position and velocity in each iteration based on Equations (7) and (8):

5
 The International Journal of Multimedia & Its Applications (IJMA) Vol.14, No. 5, October 2022

(6)

(7)

(8)

where is the particle position, is the particle velocity, is the best-remembered position,
and are cognitive and social parameters, r1 and r2 are random numbers between 0 and 1.

Figure 2 shows the PSO algorithm for calculating each parent class’s approximation and detail
coefficients in the GPCR protein hierarchy.

Step 1: Assign two random positions value to each particle. The first position resembles the
wavelet family, and the second position is the decomposition level.
Step 2: Generate the wavelet approximation and detail coefficients based on the selected
wavelet family and decomposition level.
Step 3: Evaluate the classification accuracy for each particle using the fitness function in
Equation (6).
Step 4: Compare the classification accuracy for each particle to the accuracy from the pbest
position. If the particle’s classification accuracy value is better than the pbest value,
update the pbest value using the latest classification accuracy value.
Step 5: Identify the particle with the highest classification accuracy value and assign as gbest.
Step 6: Update the position and velocity of each particle using Equations (7) and (8).
Step 7: Repeat Steps 2 to 5 until the number of iterations does not exceed the maximum
iteration or has found a high accuracy value.

Figure 2. Pseudo code of the Particle Swarm Optimisation Algorithm

The parameter values for the PSO algorithm set in this study are presented in Table 1. The
parameters and are based on the parameters set in [50].

Table 1. The parameter list for the PSO algorithm

Parameter Value
Particle size 30
Number of iteration 100
, 1.49445

6
 The International Journal of Multimedia & Its Applications (IJMA) Vol.14, No. 5, October 2022
The coefficients or features generated from the selected wavelet family and decomposition level
that produced the highest accuracy for each class are selected as features representing the class in
the hierarchy. The next step is to classify the features in the hierarchical structure.

2.3. Hierarchical Classification

In this study, the type of hierarchy used is a tree, and the type of hierarchical classification used is
Local Classification at the Parent Node (LCPN). Figure 3 shows an example of a tree hierarchy
consisting of three levels. This hierarchy has five parent nodes: the root node, nodes 1, 2, 2.1, and
2.2. A leaf or terminal node consists of nodes 1.1, 1.2, 2.1.1, 2.1.2, 2.2.1, and 2.2.2. The
classification model is trained using the exclusive sibling policy, as described in [52]. For example,
in Figure 3, the shaded box indicates node 2.1. The positive data set for node 2.1 is node 2.1, while
the negative data set for node 2.1 is node 2.2.

Figure 3. Hierarchical classification using Local Classification on the Parent Node (LCPN)

Local classification methods are divided into three types: local classification at each node (LCN),
local classification at each hierarchical level (LCL), and local classification at each parent node
(LCPN). Different methods are used to train hierarchical classification depending on the type of
local classification. In this study, the support vector machine (SVM) classifier is located on each
parent node except the leaf node. The SVM classifier performs well using the PseAAC feature
representation method [53]. Therefore, in this study, the classifier used in each parent node in the
hierarchy is the SVM classifier. Each SVM model’s training is done using a positive training dataset
from the parent node and a negative dataset from the parent’s sibling nodes.

The top-down hierarchical classification method is used to classify the test data for the classification
phase. Referring to Figure 3, the classifier in the root node will classify the data to the node in the
first level, which is class 1 or 2 only. This data will then be sent to the second level according to
the classification results from the root node. For example, if the SVM classifier classifies new data
into class 2, this data will be classified by the classifier into class 2.1 or 2.2. If the data is classified
as node 2.1, the next classification is to the leaf nodes, which are nodes 2.1.1 or 2.1.2. The SVM
classifier is used to predict GPCR since it is a robust classifier in multiple areas of biological
analysis [44]. This study uses the binary SVM classifier and the ECOC model available in
MATLAB R2019 a software. The coding method used in this study is one-to-one classification
(OVO), while the type of kernel used is Radial Basis Function (RBF).

7
 The International Journal of Multimedia & Its Applications (IJMA) Vol.14, No. 5, October 2022
2.4. Performance Evaluation

The five-fold cross-validation method is used for the evaluation of the proposed algorithm. Data is
divided into five subsets that have almost the same size. After training the SVM classifier using
four subsets, the classifier’s performance is tested using the fifth subset. This process is repeated
five times so that each subset can be used as test data.

The effectiveness of the proposed techniques is tested based on four performance testing criteria:
accuracy (A), precision (P), recall (R), and F-score (F). These measurement values are used to
evaluate the output classification performance resulting from the developed method. Accuracy,
precision, recall, and F-score values are defined in Equations (9), (10), (11), and Equation (12).

Equation (9) is used to obtain the accuracy value for each class in the hierarchy, where c(i) is the
number of protein sequences correctly classified in class i and Total (i) is the total number of protein
sequences in class i. In Equations (10), (11), and (12), the values of true positive(TP), false negative
(FN), true negative (TN), and false positive (FP) have been used. A true positive (TP) is a situation
when a positive case is successfully classified as a positive class, and a false negative (FN) is a
positive case classified as a negative case. A true negative (TN) is a negative case successfully
classified as negative, and a false positive (FP) is a case that is negative but classified as positive.

3. RESULTS AND DISCUSSIONS

This paper used the optimised DWT features and SVM classifier in the GPCR protein hierarchical
classification model. The results were analysed using three hierarchy levels: family, subfamily, and
sub-subfamily. The family, subfamily, and sub-subfamily levels consist of 5, 38, and 87 classes,
respectively.

Figure 4(a) shows the PseAAC features vector graph for one of the class A GPCR proteins. The
DWT transformed the PseAAC features into approximation and detail coefficients using the Coiflet
4 wavelet family and three decomposition levels as an example. Graphs 4(b), 4(c), and 4(d) show
the detailed wavelet coefficients for decomposition levels 1, 2, and 3 containing local features of
the protein. In contrast, graph (d) shows the wavelet approximation coefficient with the protein’s
global features or rough characteristics.

For further discussion, the results are divided into two parts. The first part discusses the selected
wavelet family and decomposition level for each parent node in the GPCR protein hierarchy, and
the second part discusses the proposed algorithm performances.

8
 The International Journal of Multimedia & Its Applications (IJMA) Vol.14, No. 5, October 2022

Figure 4. Representation of PseAAC features for class A GPCR proteins using the Coiflet4 wavelet family
with three decomposition levels

3.1. Selection of Wavelet Family and Decomposition Level

Table 2 shows the types of wavelet families selected by the PSO algorithm for five folds of the
dataset. Each data set fold contained 11 wavelet families to represent 11 parent node classes in the
GPCR protein hierarchy, which are the root, family A, family B, family C, subfamily A Amine,
subfamily A Hormone, subfamily A Nucleotide, subfamily A Peptide, subfamily A Thyro,
subfamily Prostanoid, and subfamily C CalcSense. Therefore, the total amount wavelet families for
the five folds of the data set are 55 wavelet families. The wavelet family and decomposition level
selected bythe PSO method at eleven hierarchy nodes are analysed. It is clear from Table 2 that the
wavelet family and decomposition level selected at different parent nodes differ considerably. It
corroborates the use of PSO, which automatically determines each class node’s wavelet family and
decomposition level.

Table 2. Selected wavelet family and decomposition level at each GPCR node

Class node Wavelet family and decomposition level

Root (RBior 3.7,1), (Bior4.4,3), (Bior5.5,3), (Bior2.2,3), (Bior3.7,1)
Family A (RBior4.4,2), (FK14,3), (Db1,1), (Db8,2), (Db3,1)
Family B (RBior4.4,2), (RBior3.3,4), (Db3,1), (Bior2.8,2), (Sym5,2)
Family C (Db8,2), (Db1,1), (Bior5.5,2), (FK4,2), (Rbio1.4,2)
Subfamily A A mine (Bior3.7,1), (Sym10,3), (Db1,1), (Bior1.5,4), (RBio3.3,2)
Subfamily A Hormone (Bior2.8,1), (FK4,5), (Bior4.4,5), (Bior2.2,7), (RBio2.2,6)
Subfamily A Nucleotide (Sym10,7), (Sym5,6), (Sym8,1), (Sym10,2), (Db8,3)
Subfamily A Peptide (RBio6.8,1), (Bior2.2,1), (Db1,4), (Rbio1.1,3), (Db8,1)
Subfamily A Thyro (Bior3.9,6), (Bior3.1,2), (RBio3.7,5), (Sym9,6), (Bior5.5,5)
Subfamily A Prostanoid (Bior1.5,3), (Bior4.4,2), (RBio1.5,2), (Sym8,6), (RBio2.2,6)
Subfamily C CalcSense (Bior3.3,5), (Bior2.2,4), (RBio1.5,2), (Sym5,2), (Rbio1.3,7)

9
 The International Journal of Multimedia & Its Applications (IJMA) Vol.14, No. 5, October 2022

Figure 5. Selection frequency of each wavelet family at each node of the GPCR hierarchy

Figure 5 shows the selection frequency of the wavelet family at each parent node of the GPCR
hierarchy. Out of 55 wavelet families, 19 wavelet families are selected from the Biorthogonal (Bior)
family, 14 from the family Reverse Biorthogonal (Rbio), ten from the Daubechies (Db) family, nine
from the Symlet(Sym) family, and three from the Fejer-Korovkin (FK) family. The Biorthogonal
wavelet family is the choice for representing most protein classes because most scaling and wavelet
functions in this family have a sudden shape change. It is consistent with PseAAC features as in
Figure 4(a), for example, which have a rough shape and many sharp variations. Additionally,
according to [54], the Biorthogonal wavelet can eliminate redundant information from protein
sequences, minimise feature information leakage and aliasing, allow feature vectors to represent
the original sequence information, and enhance prediction performance. In addition, PSO selects
Daubechies and Fejer-Korovkin wavelets at least nine times. In contrast, the Coiflet wavelet family
is not an option to express the PseAAC feature because the scaling and wavelet functions did not
resemble the PseAAC feature in this study.

Figure 6 shows the frequency of selected decomposition levels to represent protein classes.
Decomposition level 1 produced half of the 170 PseAAC features as 85 approximation and 85 detail
coefficients. Therefore, decomposition level 1 allowed the preservation of global and local
information. Decomposition level 2 produced half of the 85 PseAAC features as 42 approximation
and 42 detail coefficients. This process continues for the rest of the decomposition levels. Out of
55 wavelet families, 16 used decomposition level 2, the highest decomposition level used in feature
representation. It is followed by 13 wavelet families using decomposition level 1 and eight wavelet
families using decomposition level 3. Four wavelet families used decomposition level 4, five
families chose decomposition level 5, and six families chose decomposition level 6. The least
decomposition level selected for the feature representation is decomposition level 7, which is 3.
The greater decomposition level will increase the number of detail coefficients and decrease the
number of approximation coefficients. The protein class requires more local information than
global information to represent it.

10
 The International Journal of Multimedia & Its Applications (IJMA) Vol.14, No. 5, October 2022

Figure 6. Selection frequency of each decomposition level at each node of the GPCR hierarchy

3.2. GPCR Classification Performance Without Virtual Class

Table 3 shows the performance for classifying GPCR using SVM at the family level. It can be
observed from the results show that, at the family level, DWT+PSO is the best feature extraction
strategy. The performance achieved by DWT+PSO is 97.9% for accuracy, precision, and recall,
and the F-score achieved is 0.979.

Table 4 shows the performance for classifying GPCR using SVM at the subfamily level. As shown
in the table, DWT+PSO is the best feature extraction strategy at the subfamily level at this hierarchy
level. The performance achieved by DWT+PSO is 85.9%, 88.6.7%, 88.4%, and 0.885% for
accuracy, precision, recall, and F-score, respectively. Nevertheless, the DWT+PSO algorithm
achieved almost the same accuracy as PseAAC, which is 85%.

Table 3. GPCR classification performance for family level

Feature
extraction Accuracy (%) Precision (%) Recall (%) F-score
method
PseAAC 97.7 97.7 97.7 0.977
DWT+PSO 97.9 97.9 97.9 0.979

Table 5 shows the performance for classifying GPCR using SVM at the sub-subfamily level. It can
be seen from this table that, at the sub-subfamily level, the best accuracy comes from DWT+PSO,
which is 77.5%. The best precision, recall, and F-score come from DWT+PSO, 87.7%, 82.3%, and
0.849, respectively.

Tables 4 and 5 show that the precision value is higher than the accuracy value for the classification
results using PseAAC and PSO at the subfamily and sub-subfamily levels. However, the accuracy
value is not a good metric for imbalanced data sets, and it is because high accuracy can also be
achieved by successfully classifying the dominant negative class. Since it has been confirmed that
the GPCR protein has imbalanced data for each class ([55], [40]), getting a high precision value is
more important ([56]). It is the same case in Table 5, whereby the recall value is higher than the
accuracy value. It also denotes that DWT+PSO performed well in an imbalanced GPCR protein
dataset.

11
 The International Journal of Multimedia & Its Applications (IJMA) Vol.14, No. 5, October 2022
Table 4. GPCR classification performance for sub family level

Feature extraction
method Accuracy (%) Precision (%) Recall (%) F-score

PseAAC 85.3 87.7 87.7 0.877

DWT+PSO 85.9 88.6 88.4 0.885

Table 5. GPCR classification performance for sub-sub family level

Feature extraction
method Accuracy (%) Precision (%) Recall (%) F-score

PseAAC 76.9 86.2 77.4 0.816

DWT+PSO 77.5 87.7 82.3 0.849

Table 6 compares GPCR protein classification accuracy performance between previous studies
using the GDS data set and this study. It was found that this study produced an accuracy value of
97.9% at the family level, which is almost the same as the accuracy value in the study of [42], [40],
[41], [57], and [39]. This study obtained an accuracy value of 86% for the second level of the
hierarchy. However, [40] has obtained higher accuracy values for the second level of the hierarchy,
which is 89.2%. For the third level, the accuracy value obtained from this study is 78.3%. The
highest accuracy value for the third level was obtained from the study of [40], with an accuracy
value of 90.95%.

Table 6. Comparison of GPCR protein classification using the GDS dataset

Authors Super Family Subfamily Sub-subfamily

Family
[37] - 90.59% 73.77% 58.08%
[43] - 96.97% 82.72% 70.46%
[42] 99.75% 97.38% 81.91% 73.34%
[41] - 97.41% 84.97% 75.60%
[40] 98% 97.5% 89.2% 90.95%
[57] - 97.40% 87.78% 81.13%
[39] - 97.17% 86.82% 81.17%
This study - 97.9% 86% 78.3%

Although the accuracy of this study is lower than [40] at the subfamily and sub-subfamily level,
their research focused on optimising the GPCR protein classifiers. In contrast, this study tried to
optimise the representation of GPCR proteins at the feature level. Hence, compared to other studies,
it was found that this study’s results are comparable to [57] and [39] at the family and subfamily
levels, which used the current trending deep learning method.

4. CONCLUSIONS
In this study, the particle swarm optimisation (PSO) algorithm has helped to identify the wavelet
family and the appropriate decomposition level for each parent class representation in the GPCR

12
 The International Journal of Multimedia & Its Applications (IJMA) Vol.14, No. 5, October 2022
protein hierarchy. There is still much room for improvement and development in this study. The
imbalance data problem in the data set is always encountered in data mining. The GPCR dataset
also suffers from the same problem in which 60% of the known GPCR protein sequences are
proteins from the A family. The sampling process must be considered to overcome this problem,
such as using the Synthetic Minority Over sampling Technique (SMOTE) algorithm. SMOTE is an
oversampling technique generating synthetic samples from minority classes, and it is used to obtain
a synthetically balanced training set for each class and then train the classifier.

Future research is welcome to study the appropriate wavelet family and decomposition levels for
GPCR protein class representation to perform better. The experiments conducted in this study
utilised parameter values for the PSO algorithm based on previous studies. This study did not
examine whether the parameters are appropriate values for PSO. Therefore, there is a need for a
comprehensive study to determine the appropriate parameter values required for the PSO algorithm
as an algorithm for selecting the wavelet family and the decomposition level.

There are several approaches to hierarchical classification that can be implemented. It includes the
global hierarchical classification method that does not suffer the propagated errors problem. This
method generates a complex classification model from a training set for global classification
andconsiders the class hierarchy as a whole during classification. Each test data will be classified
based on the classification model produced for the test phase. It is hoped that this global hierarchical
classification method can help improve the hierarchical classification performance of GPCR
proteins.

ACKNOWLEDGEMENT

This study was financially supported by the Malaysia Ministry of Higher Education (MOHE).

REFERENCES

[1] K. Alhosaini, A. Azhar, A. Alonazi, and F. Al-Zoghaibi, "GPCRs: The most promiscuous druggable
receptor of the mankind," Saudi Pharm. J., no. May, 2021, doi: 10.1016/j.jsps.2021.04.015.
[2] M. Li, C. Ling, and J. Gao, "An Efficient CNN-based Classification on G-protein Coupled Receptors
Using TF-IDF and N-gram," 2017 IEEE Symp. Comput. Commun., pp. 924–931, 2017.
[3] M. Davies, A. Secker, and A. Freitas, "Optimising amino acid groupings for GPCR classification,
"Bioinformatics, vol. 24, no. 18, pp. 1980–1986, 2008, doi: 10.1093/bioinformatics/btn382.
[4] R. Karchin, K. Karplus, and D. Haussler, "Classifying G-protein coupled receptors with support vector
machines,"Bioinformatics, vol. 18, no. 1, pp. 147–159, 2002, doi: 10.1093/bioinformatics/18.1.147.
[5] S. Saini and L. Dewan, "Comparison of Numerical Representations of Genomic Sequences: Choosing
the Best Mapping for Wavelet Analysis, "Int. J. Appl. Comput. Math., vol.3, no.4, pp. 2943–2958, 2017,
doi: 10.1007/s40819-016-0277-1.
[6] T. T. Gayathri and S. A. Christe, "Wavelet Analysis in Prediction and Identification of Cancerous
Genes, "Int. J. Sci. Eng. Res., vol. 8, no. 3, pp. 720–727, 2017.
[7] W. Hou, Q.Pan, Q. Peng, and M. He, "A new method to analyse protein sequence similarity using
Dynamic Time Warping, "Genomics J., vol. 109, no. 2, pp. 123–130, 2017.
[8] T. Mengetal., "Wavelet analysis in current cancer genome research: A survey, "IEEE/ACM Trans.
Comput. Biol. Bioinforma., vol. 10, no. 6, pp. 1442–1459, 2013, doi: 10.1109/TCBB.2013.134.
[9] J.-D. Qiu, X.-Y. Sun, J.-H. Huang, and R.-P. Liang, "Prediction of the types of membrane proteins
based on discrete wavelet transform and support vector machines., "Protein J., vol. 29, no. 2, pp. 114–
9, 2010, doi: 10.1007/s10930-010-9230-z.
[10] B. Chen, Y. Li, and N. Zeng, "Centralized Wavelet Multiresolution for Exact Translation Invariant
Processing of ECG Signals, "IEEE Access, vol. 7, pp. 42322–42330, 2019, doi:
10.1109/ACCESS.2019.2907249.
[11] A. Elbir, H. O. Ilhan, G. Serbes, and N. Aydin, "Short Time Fourier Transform based music genre

13
 The International Journal of Multimedia & Its Applications (IJMA) Vol.14, No. 5, October 2022
classification,"2018 Electr. Electron. Comput. Sci. Biomed. Eng. Meet. EBBT2018, no. June, pp. 1– 4,
2018, doi: 10.1109/EBBT.2018.8391437.
[12] C. C. Aggarwal, "On effective classification of strings with wavelets, " Proc. ACM SIGKDD Int. Conf.
Knowl. Discov. Data Min., pp. 163–172, 2002, doi: 10.1145/775069.775071.
[13] T. D. Mai, T .D. Ngo, D. D. Le, D .A. Duong, K. Hoang, and S. Satoh, "Using node relationships for
hierarchical classification, "Proc. - Int. Conf. Image Process. ICIP, vol. 2016-Augus, pp. 514–518,
2016, doi: 10.1109/ICIP.2016.7532410.
[14] B. Yu et al., "Prediction of protein structural class for low-similarity sequences using Chou's pseudo
amino acid composition and wavelet denoising,"J. Mol. Graph. Model., vol. 76, no. July, pp. 260– 273,
2017, doi: 10.1016/j.jmgm.2017.07.012.
[15] C. de Trad, Q. Fang, and I. Cosic, "An overview of protein sequence comparisons using wavelets,
"Proc. IEEE-EMBS, 2001, Accessed: Mar. 29, 2014. [Online]. Available:
http://www.eng.monash.edu/non-cms/ecse/ieee/ieeebio2001/trad.pdf.
[16] P. Liò, "Wavelets in bioinformatics and computational biology: State of art and perspectives,"
Bioinformatics, vol. 19, no.1, pp. 2–9, 2003, doi:10.1093/bioinformatics/19.1.2.
[17] A. D. Haimovich, B. Byrne, R. Ramaswamy, and W. J. Welsh, "Wavelet analysis of DNA walks, "J.
Comput. Biol., vol. 13, no. 7, pp. 1289–1298, 2006, doi: 10.1089/cmb.2006.13.1289.
[18] Z. Germán-Salló and G. Strnad, "Signal processing methods in fault detection in manufacturing
systems, "Procedia Manuf., vol. 22, pp. 613–620, 2018, doi: 10.1016/j.promfg.2018.03.089.
[19] J.-D. Qiu, J.-H. Huang, R.-P. Liang, and X.-Q. Lu, "Prediction of G-protein-coupled receptor classes
based on the concept of Chou's pseudo amino acid composition: an approach from discrete wavelet
transform., "Anal. Biochem., vol. 390, no. 1, pp. 68–73, Jul. 2009, doi: 10.1016/j.ab.2009.04.009.
[20] Y. I. Jang, J. Y. Sim, J. R. Yang, and N. K. K won, "The optimal selection of mother wavelet function
and decomposition level for denoising of dcg signal, "Sensors, vol. 21, no. 5, pp. 1–17, 2021, doi:
10.3390/s21051851.
[21] S. Saini and L. Dewan, "Performance comparison of first generation and second generation wavelets
in the perspective of genomic sequence analysis, "Int. J. Pure Appl. Math., vol. 118, no. 16, pp. 417–
442, 2018.
[22] H. He, Y. Tan, and Y. Wang, "Optimal base wavelet selection for ECG noise reduction using a
comprehensive entropy criterion, "Entropy, vol. 17, no. 9, pp. 6093–6109, 2015, doi:
10.3390/e17096093.
[23] W. K. Ngui, M. S. Leong, L. M. Hee, and A. M. Abdelrhman, "Wavelet analysis: Mother wavelet
selection methods, "Appl. Mech. Mater., vol. 393, no. January 2014, pp. 953–958, 2013, doi:
10.4028/www.scientific.net/AMM.393.953.
[24] T. Wang, L. Li, Y. A. Huang, H. Zhang, Y. Ma, and X. Zhou, "Prediction of protein-protein interactions
from amino acid sequences based on continuous and discrete wavelet transform features, "Molecules,
vol. 23, no. 4, 2018, doi: 10.3390/molecules23040823.
[25] D. Chen, S. Wan, J. Xiang, and F. S. Bao, "A high-performance seizure detection algorithm based on
Discrete Wavelet Transform (DWT) and EEG, "PLoS One, vol. 12, no. 3, Mar. 2017, doi:
10.1371/journal.pone.0173138.
[26] C. Guarnizo, a a Orozco, and M. a Alvarez, "Optimal sampling frequency in wavelet-based signal
feature extraction using particle swarm optimisation., "Conf. Proc. IEEE Eng. Med. Biol .Soc., vol.
2013, pp. 993–6, Jan. 2013, doi: 10.1109/EMBC.2013.6609670.
[27] J.-D. Qiu, J.-H. Huang, S.-P. Shi, and R.-P. Liang, "Using the Concept of Chous Pseudo Amino Acid
Composition to Predict Enzyme Family Classes: An Approach with Support Vector Machine Based on
Discrete Wavelet Transform, " Protein Pept. Lett., vol. 17, no. 6, pp. 715–722, 2010, doi:
10.2174/092986610791190372.
[28] F. M. Albkosh, M. S. Hitam, W. N. J. H. Wan Yussof, A. A. K. Abdul Hamid, and R. Ali, "Optimisation
of discrete wavelet transform features using artificial bee colony algorithm for texture image
classification, "Int. J. Electr. Comput. Eng., vol. 9, no. 6, pp. 5253–5262, 2019, doi:
10.11591/ijece.v9i6.pp5253-5262.
[29] C. Caramia, C. De Marchis, and M. Schmid, "Optimising the scale of a wavelet-based method for the
detection of gait events from a waist-mounted accelerometer under different walking speeds, "Sensors
(Switzerland), vol. 19, no. 8, 2019, doi: 10.3390/s19081869.
[30] Z. Zhang, Q. K. Telesford, C. Giusti, K. O. Lim, and D. S. Bassett, "Choosing wavelet methods, filters,
and lengths for functional brain network construction, " PLoS One, vol. 11, no. 6, pp. 1–24, 2016, doi:
10.1371/journal.pone.0157243.

14
 The International Journal of Multimedia & Its Applications (IJMA) Vol.14, No. 5, October 2022
[31] N. Ahuja, L. Lertrattanapanich, and N. K. Bose, "Properties determining choice of mother wavelet,
"IEE proceedings. Vision, image signal Process., vol. 152, no. 5, pp. 205–212, 2005, doi: 10.1049/ip-
vis.
[32] Z. A. A. Alyasseri, A. T. Khader, M. A. Al-Betar, A. K. Abasi, and S. N. Makhadmeh, "EEG Signals
Denoising Using Optimal Wavelet Transform Hybridized with Efficient Metaheuristic Methods, "IEEE
Access, vol. 8, pp. 10584–10605, 2020, doi: 10.1109/ACCESS.2019.2962658.
[33] G. Oltean and L. N. Ivanciu, "Computational intelligence and wavelet transform based metamodel for
efficient generation of not-yet simulated waveforms, " PLoS One, vol. 11, no. 1, pp. 1–30, 2016, doi:
10.1371/journal.pone.0146602.
[34] H. Tao, J. M. Zain, M. M. Ahmed, A. N. Abdalla, and W. Jing, "A wavelet-based particle swarm
optimisation algorithm for digital image watermarking, "Integr. Comput. Aided. Eng., vol.19, no.1,
pp. 81–91, 2012, doi:10.3233/ICA-2012-0392.
[35] H. Aprillia, H. T. Yang, and C. M. Huang, "Optimal decomposition and reconstruction of discrete
wavelet transformation for short-term load forecasting, "Energies, vol. 12, no. 24, 2019, doi:
10.3390/en12244654.
[36] A. Semnani, L. Wang, M. Ostadhassan, M. Nabi-Bidhendi, and B. N. Araabi, "Time-frequency
decomposition of seismic signals via quantum swarm evolutionary matching pursuit, "Geophys.
Prospect., vol. 67, no. 7, pp. 1701–1719, 2019, doi: 10.1111/1365-2478.12767.
[37] M. N. Davies, A. Secker, A. a Freitas, M. Mendao, J. Timmis, and D. R . Flower, "On the hierarchical
classification of G protein-coupled receptors., "Bioinformatics, vol. 23, no. 23, pp. 3113–8, Dec. 2007,
doi: 10.1093/bioinformatics/btm506.
[38] A. Secker, M. N. Davies, A. A. Freitas, J. Timmis, M. Mendao, and D. R. Flower, "An Experimental
Comparison of Classification Algorithms for the Hierarchical Prediction of Protein Function
Classification of GPCRs, "Proc. 3rd UK Data Min. Knowl. Discov. Symp., 2007.
[39] S. Seo, M. Oh, Y. Park, and S. Kim, "DeepFam: Deep learning based alignment-free method for protein
family modeling and prediction, " Bioinformatics, vol. 34, no. 13, pp. i254–i262, 2018, doi:
10.1093/bioinformatics/bty275.
[40] M. Zekri, K. Alem, and L. Souici-Meslati, "Immunological Computation for Protein Function
Prediction, "Fundam. Informaticae, vol. 139, no. February 2014, pp. 91–114, 2015, doi: 10.3233/FI-
2015-1227.
[41] Z.-U. Rehman, M. T. Mirza, A. Khan, and H. Xhaard, "Predicting G-protein-coupled receptors families
using different physio chemical properties and pseudo amino acid composition., "Methods Enzymol.,
vol. 522, pp. 61–79, Jan. 2013, doi: 10.1016/B978-0-12-407865-9.00004-2.
[42] M. Naveed and A. U. Khan, "GPCR -MPredictor: Multi-level prediction of G protein-coupled receptors
using genetic ensemble, "Amino Acids, vol. 42, no. 5, pp. 1809–1823, 2012, doi: 10.1007/s00726-011-
0902-6.
[43] A. Secker, M. N. Davies, A. A. Freitas, J. Timmis, E. Clark, and D. R. Flower, "An artificial immune
system for clustering amino acids in the context of protein function classification,"J. Math. Model.
Algorithms, vol. 8, no. 2, pp. 103–123, 2009, doi: 10.1007/s10852-009-9107-3.
[44] Q. Bin Gao, X. F. Ye, and J. He, "Classifying G-protein–coupled receptors to the finest subtype level,
"Biochem. Biophys. Res. Commun., vol. 439, no. 2, pp. 303–308, 2013, doi: 10.1016/j.bbrc.2013.08.023.
[45] H. Bin Shen and K. C. Chou, "PseAAC: A flexible web server for generating various kinds of protein
pseudo amino acid composition, "Anal. Biochem., vol. 373, no. 2, pp. 386–388, 2008, doi:
10.1016/j.ab.2007.10.012.
[46] F. Y. Dao et al, "Recent advances in conotoxin classification by using machine learning methods,"
Molecules, vol. 22, no. 7, 2017, doi: 10.3390/molecules22071057.
[47] A. Shaker, "Comparison Between Orthogonal and Bi-Orthogonal Wavelets, "J. Southwest Jiatong
Univ., vol. 55, no. 2, 2020.
[48] A. Dogra, B. Goyal, and S .Agrawal, "Performance Comparison of Different, "Asian J. Pharm., vol.
2016, no. 4, pp. 9–12, 2016.
[49] J. Kennedy and R. Eberhart, "Particle Swarm Optimisation, "Proc. IEEE Int. Conf. Neural Networks,
pp. 1942–1948, 1995, doi: 10.1007/978-3-030-61111-8_2.
[50] İ. B. Aydilek, "A hybrid firefly and particle swarm optimisation algorithm for computationally
expensive numerical problems, "Appl. Soft Comput. J., vol. 66, no. February 2018, pp. 232–249, 2018,
doi: 10.1016/j.asoc.2018.02.025.
[51] T. T. Ngo, A. Sadollah, and J. H. Kim, "A cooperative particle swarm optimiser with stochastic
movements for computationally expensive numerical optimisation problems, "J. Comput. Sci., vol. 13,
pp. 68–82, 2016, doi: 10.1016/j.jocs.2016.01.004.
15
 The International Journal of Multimedia & Its Applications (IJMA) Vol.14, No. 5, October 2022
[52] C. N. Silla and A. A. Freitas, "Selecting different protein representations and classification algorithms
in hierarchical protein function prediction, "Intell. Data Anal. Journal. Vol. 15, No. 6, vol. 44, no. 0, pp.
979–999, 2011.
[53] S. Bekhouche and Y. M. Ben Ali, "Optimising the identification of GPCR function, " ACMInt. Conf.
Proceeding Ser., 2019, doi: 10.1145/3314074.3314082.
[54] B. Yu et al., "Prediction subcellular localisation of Gram-negative bacterial proteins by support vector
machine using wavelet denoising and Chou's pseudo amino acid composition,"Chemom. Intell. Lab.
Syst., vol. 167, no. October, pp. 102–112, 2017, doi: 10.1016/j.chemolab.2017.05.009.
[55] Q. Gu, Y.-S. Ding, and T.-L. Zhang, "Prediction of G-Protein-Coupled Receptor Classes in Low
Homology Using Chous Pseudo Amino Acid Composition with Approximate Entropy and
Hydrophobicity Patterns, "Protein Pept. Lett., vol. 17, no. 5, pp. 559–567, 2010, doi:
10.2174/092986610791112693.
[56] B. Juba and H. S. Le, "Precision-Recall versus accuracy and the role of large data sets, "33rd AAAI
Conf. Artif. Intell. AAAI 2019, 31st Innov. Appl. Artif. Intell. Conf. IAAI 2019 9th AAAI Symp. Educ.
Adv. Artif. Intell. EAAI 2019, pp. 4039–4048, 2019, doi: 10.1609/aaai.v33i01.33014039.
[57] R. Paki, E. Nourani, and D. Farajzadeh, "Classification of G protein–coupled receptors using attention
mechanism," Gene Reports, vol. 21, no. August, p. 100882, 2020, doi: 10.1016/j.genrep.2020.100882.

16