MR.

MUHAMMED TILAHUN MUHAMMED (Orcid ID : 0000-0003-0050-5271)

Accepted Article
Article type : Review

Homology Modeling in Drug Discovery: Overview, Current Applications and Future Perspectives

Muhammed Tilahun Muhammed*1,2, Esin Aki-Yalcin3

1
Department of Pharmaceutical Chemistry, Faculty of Pharmacy, Suleyman Demirel University,
32260 Isparta, Turkey.
2
Department of Basic Biotechnology, Institute of Biotechnology, Ankara University, Tandogan, 06100
Ankara, Turkey.
3
Department of Pharmaceutical Chemistry, Faculty of Pharmacy, Ankara University, Tandogan, 06100
Ankara, Turkey; Esin.Aki@ankara.edu.tr
*
Correspondence: muh.tila@gmail.com; Tel.: +90-246-211-0342

Abstract

Homology modeling is one of the computational structure prediction methods that are used to
determine protein 3D structure from its amino acid sequence. It is considered to be the most
accurate of the computational structure prediction methods. It consists of multiple steps that are
straightforward and easy to apply.

There are many tools and servers that are used for homology modeling. There is no single modeling
program or server which is superior in every aspect to others. Since the functionality of the model
depends on the quality of the generated protein 3D structure, maximizing the quality of homology
modeling is crucial.

Homology modeling has many applications in the drug discovery process. Since drugs interact with
receptors, which consists mainly of proteins in their structure, protein 3D structure determination,
and thus homology modeling is important in drug discovery. Accordingly, there has been the
clarification of protein interactions using 3D structures of proteins that are built with homology
modeling. This contributes to the identification of novel drug candidates.

Homology modeling plays an important role in making drug discovery faster, easier, cheaper and
more practical. As new modeling methods and combinations are introduced, the scope of its
applications widens.

This article has been accepted for publication and undergone full peer review but has not been
through the copyediting, typesetting, pagination and proofreading process, which may lead to
differences between this version and the Version of Record. Please cite this article as doi:
10.1111/cbdd.13388
This article is protected by copyright. All rights reserved.
 Keywords: Current application; Drug discovery; Homology modeling; Structure prediction; 3D
structure
Accepted Article
Introduction

The world wide Protein Data Bank (wwPDB) (https://www.wwpdb.org/) contains approximately 144,
000 experimentally determined protein three dimensional (3D) structures currently [1]. In contrast
the last reference sequence, which is a non redundant sequence, release of National Center for
Biotechnology Information (NCBI) (https://www.ncbi.nlm.nih.gov/) consists of annotated 155 million
sequences including approximately 106 million protein sequences [2]. This represents a protein
sequence number that is 736 times larger than the protein 3D strucuture deposited in the wwPDB. In
2006 the annotated sequence in NCBI was nearly 120 times larger than experimentally solved 3D
structures deposited in wwPDB [3]. This means the number of protein sequences has increased 6
times faster than the number of experimentally determined protein 3D structures. Since the protein
data banks available contain redundancy but the sequences in NCBI are non redundant, the
difference is higher than the numbers given here. This growing gap between the sequences available
and the protein 3D structures determined is in an alarming condition. Thus, computational structural
determination methods are needed in filling this widening gap between the number of sequences
available and protein 3D structures solved experimentally.

Since crystal structure of the first protein myoglobin was solved in 1960, there has been an
improvement in the quality of the 3D structures determined. This has been achieved with the
introduction of experimental methods like X-ray crystallography and NMR spectroscopy [4].
However, these experimental methods can not be used for each protein. For NMR analysis protein
molecules should be small and for X-ray crystallography the molecules should be crystallized.
Additionally, these methods are time consuming. Thus, there is deficiency in high resolution 3D
structure of proteins, especially membrane proteins due to the difficulties in purification and
crystallization of such proteins in relative to other small water soluble proteins [5]. Since membrane
proteins constitute important proportion of therapeutic drug targets, advances in the determination
of membrane proteins will speed up the drug discovery process. Here computational protein 3D
structure prediction can play a crucial role.

Homology modeling (comparative modeling) is one of the computational structure prediction

methods that are used to determine 3D structure of a protein from its amino acid sequence based on
its template. The basis for homology modeling are two major observations. First protein 3D structure
is particularly determined by its amino acid sequence. Second the structure of proteins is more
conserved and the change happens at a much slower rate in relative to the sequence during
evolution. As a result, similar sequences fold into identical structures and even sequences with low
relation take similar structures [6].

Homology modeling is considered to be the most accurate of the computational structure prediction
methods [7]. 3D structure predictions made by computational methods like de novo prediction and
threading were compared to homology modeling using Root Mean Square Deviation (RMSD) as a
criteria. Homology modeling was found to give 3D structures with the highest accuracy [8].
Furthermore, it is a protein 3D structure prediction method that needs less time and lower cost with

This article is protected by copyright. All rights reserved.

 clear steps. Thus, homology modeling is widely used for the generation of 3D structures of proteins
with high quality. This has changed the ways of docking and virtual screening methods that are based
Accepted Article on structure in the drug discovery process [9].

In this review, the main features of steps of homology modeling are presented. The popular tools and
servers that have been used for homology modeling in recent years are also summarized. Overview
of the striking homology modeling applications in the prediction of protein 3D structures and recent
applications in the drug discovery are also discussed. This review also provides insight into the
opportunities and possible challenges in homology modeling.

Steps of Homology Modeling

Homology modeling is a structure prediction method that consists of multiple steps. Homology
modeling has common standard procedures with minor differences. The standard steps of homology
modeling are summarized in Figure 1 and the detail explanation is given below the figure.

Identification and Selection of Templates

In this step of the process target (query) sequence is used for the identification of template
structures in the PDB (https://www.rcsb.org/) [10] or similar databases. There are popular tools in
searching for eligible templates for target sequence with different approaches. Among of them, Basic
Local Alignment Search Tool (BLAST) [11] is the one which provides pairwise sequence-sequence
alignment. This service is available inside databases like NCBI [2] and UniProt
(http://www.uniprot.org/) [12]. The other approaches used in template identification are profile-
profile alignments [13] and Hidden Markov Models (HMMs) [14]. Some other advanced approaches
use profile-profiles and HMMs in combination with structural properties.

After template candidates are identified, the best structures must be selected. Sequence similarity
level of the template sequence in relative to the target sequence is important in generating 3D
structures with high accuracy. However, sequence similarity is not the only factor that determines
the accuracy of the structures generated in homology modeling. Regarding the minimum sequence
similarity limit in homology modeling, there are ambiguities about the exact value but >25% suggests
that the template and target will take similar 3D structures [15].

Apart from high sequence similarity, various factors are considered in choosing an eligible template.
These factors include phylogenetic similarity between template and target sequences. Templates
from identical or analogous phylogenetic tree to the target sequence may result in a 3D structure
with high accuracy [16]. The other factors are environmental factors such as pH, solvent type and
existence of bound ligand. These are also important in choosing the most eligible template as it has a
role in ensuring the most optimal conditions in building an accurate target structure. The resolution
of the experimental structure under consideration is also a factor in choosing the eligible template
[17].

This article is protected by copyright. All rights reserved.

 Sequence Alignments and Alignment Correction
Accepted Article After the most appropriate alignments are selected, alignments and correction of them in case it is
necessary is undertaken. The alignments are target-template and template-template when more
than one template is used. The error in the alignment of a residue causes shifting of α carbon. A
single residue gap in an α helix section triggers rotation of the rest of the residues in the helix. As a
result, the alignment of sequences in the right way is crucial in homology modeling [18]. Careful
checkups and correction while performing alignments may enhance building 3D protein structures
with high quality. The most widely used alignment methods are Clustal W
(http://www.genome.jp/tools-bin/clustalw) [19], T-Coffee (http://tcoffee.crg.cat/) [20], 3Dcoffee
(http://phylogeny.lirmm.fr/phylo_cgi/) [21] and MUSCLE
(https://www.ebi.ac.uk/Tools/msa/muscle/) [22].

Model Building

Various methods are used to generate 3D models for the target sequence based on its templates.
Model building approaches can be classified as rigid body-assembly methods, segment matching
methods, spatial restraint methods and artificial evolution methods.

In rigid-body assembly the protein structure is broken down into basic conserved core regions, loops
and side chains. This approach depends on the natural dissection that enables the building of a
protein 3D structure by bringing this rigid bodies together which are picked up from the aligned
template protein structures [23]. This can be done by tools like 3D-JIGSAW [24], BUILDER [25] and
SWISS-MODEL [26].

In segment matching method a cluster of atomic positions obtained from the template structures are
used as leading positions. Selection of segments from known structures in a database for matching
the segments is done based on the sequence identity, geometry and energy. Then the entire atom
model is generated by using the leading structure as a pillar to lay the segments. This can be done by
using SEGMOD/ENCAD [27].

Spatial restraint method builds the model by meeting restraints came from the template structure.
The restraints are framed onto the target structure depending on the alignment. These restraints are
determined by stereochemical restraints on bond length, bond angle, dihedral angles and van der
waals contact distances. This can be performed with MODELLER [28].

Artificial evolution method uses rigid-body assembly method and stepwise template evolutionary
mutations together until the template sequence is the same as the target sequence. This can be
performed with NEST [29].

Table 1 displays summary of general features of the popular tools and servers that can be used for
model building. Researchers reported that when the sequence identity is high, the homology models
derived from different packages are comparable to each other. When the sequence identity is lower,
the results tend to vary, with some packages performing noticeably better than others [30]. The
quality of the models is related with the performance of packages in sequence alignment and model
building. MODELLER is found to be one of the best tools in homology modeling [31]. In addition to

This article is protected by copyright. All rights reserved.

 this critical assessment of methods of protein structure prediction (CASP) assesses modeling
methods in a number of different categories. I-TASSER was ranked as the best server for protein
Accepted Article structure prediction in recent CASP experiments [32]. These tools and servers have their own pros
and cons. As a result, there is no single modeling tool or server which is superior in every aspects to
others.

Loop Modeling

Gaps or insertions called loops are present in sequences of homologous proteins. The structures of
loops are not conserved during evolution. Even without deletions or insertions different loop
conformations in query and template are often found. The specificity of the function of a protein
structure is often determined by the loops. Accuracy of loop modeling is an important factor which
determines the value of the generated models for further applications. Since loops show higher
structural variability than strands and helices, the prediction of their structure is more difficult than
strands and helices [44].

There are two important methods that are used in developing the loops. One is database search
approach and the other is conformation search approach. The database search method browses all
the known protein structures to detect segments providing the critical core regions. The
conformational search approach depends on a scoring function optimization [8]. Loop searches are
done for loops of length 4-7 residues these days. This is because of the conformation variation
increase as the length of the loop increases.

To deal with these drawbacks, de novo methods that are used for loop conformation predictions by
looking for conformational space have been developed. Monte Carlo simulations, simulated
annealing, genetic algorithms and molecular dynamics simulations often in combination with
knowledge-based potentials are examples for this. In such methods the length of loop that can be
modelled is not limited but as the length increases possible conformation number increases rapidly
which makes the modeling very time consuming [45]. There are servers such as ArchPRED
(http://www.bioinsilico.org/ARCHPRED/) [46] and Congen
(http://www.congenomics.com/congen/doc/) [6,47] that are used in loop modeling.

Side Chain Modeling

Side chain modeling is usually done by putting side chains onto the backbone coordinates that are
derived from a parent structure and/or from ab initio modeling simulations. In practice side chain
prediction works at high levels of sequence identity. Protein side chains are present in a limited
number of structures with low energy known as rotamers. Depending on defined energy functions
and search strategies, rotamers are selected in accordance with the preferred protein sequence and
the given backbone coordinates. The accuracy of prediction is usually high for the hydrophobic core
residues but low for water exposed residues on the surface [48]. Tools like RAMP
(http://www.ram.org/computing/ramp/) [41] and SCWRL [49] can be used in side chain modeling.

This article is protected by copyright. All rights reserved.

Accepted Article Model Optimization

Optimization of the model usually begins with an energy minimization utilizing molecular mechanics
force fields [50]. At each energy minimization a few big errors are eliminated but many other small
errors are introduced at the same time and start accumulating. Therefore, restraining the atom
positions, implementing energy minimization with a few hundred steps and using more precise force
fields like quantum force fields [51] and self-parameterizing force fields [52] can be utilized to
decrease the errors in model optimization. For further model optimization methods such as
molecular dynamics and Monte Carlo can be used [53].

Model Validation

Accuracy of the constructed model can determine its further application in various areas. Thus,
verification and validation of models are necessary. Depending on sequence similarity,
environmental parameters and the quality of the templates, the generated models have different
accuracy.

Analysis of the stereochemistry of the model is one basic requirement. This analysis is done with
parameters such as bond length, torsion angle and rotational angle. WHATCHECK
(https://swift.cmbi.umcn.nl/gv/whatcheck/) [54], PROCHECK (https://www.ebi.ac.uk/thornton-
srv/software/PROCHECK/) [55] and Molprobity (http://molprobity.biochem.duke.edu/) [56] are
popular tools used for the determination of the stereochemistry of the model in homology modeling.
The Ramachandran plot (http://mordred.bioc.cam.ac.uk/~rapper/rampage.php) is also powerful
determinant of the quality of protein structure. Residues with a problem of stereochemistry will fall
out of the acceptable regions of the Ramachandran plot [57].

There are also tools that focus on the determination of the spatial features of the model based on 3D
conformations and mean force statistical potentials. VERIFY3D
(http://servicesn.mbi.ucla.edu/Verify3d/) [58] and PROSAII (https://www.came.sbg.ac.at/prosa.php)
[7] are examples for this. These tools consider model construction environmental parameters in
relative to the expected environmental conditions.

Applications of Homology Modeling

Homology modeling has a vast range of applications and its importance is increasing as the number
of structures determined increases. It has applications in structure based drug design, analysis of
mutations, insight into binding mechanisms, identification of active sites, looking for ligands and
designing of novel ligands, modeling of substrate specificity, protein–protein docking simulations,
molecular replacement in experimental structural refinements, rationalizing of known experimental
results and planning of future computational experiments by using the generated models [59].

This article is protected by copyright. All rights reserved.

 Homology modeling has many applications in drug discovery process. This makes the drug discovery
process faster, easier, cheaper and more practical. In general homology modeling applications in the
Accepted Article drug discovery need high quality models. As a result, high sequence similarity, good side chain
modeling and loop modeling are crucial in determining further applications of the model build in the
drug discovery.

As an illustration of the case application, for example, homology modeling was used to discover
novel acetohydroxy acid synthase (AHAS, EC 2.2.1.6) inhibitors against M. Tuberculosis. Several
studies demonstrated that the plant AHAS inhibitors of sulfonylurea chemicals such as sulfometuron
methyl (SMM) exhibit antituberculosis activity. However, the 3D structure of M. tuberculosis AHAS
remains to be elucidated. Thus, homology modeling was performed based on the S. cerevisiae AHAS
to build a 3D structure of M. Tuberculosis AHAS. Through docking simulation and similarity searches,
23 novel AHAS inhibitors of E. coli AHAS II enzymatic activity were identified. Five of the identified
chemicals showed strong inhibitory effects against multidrug-resistant and extensively drug-resistant
strains. Three of the compounds exhibited more activity than the positive control SMM [60].

In recent years 3D structure of targets in cancer that can be used for discovering effective
chemotherapeutic agents has been generated using homology modeling [61,62]. Reliable 3D
structures of G-protein coupled receptors (GPCRs) which are targets of nearly a third of FDA
approved drugs has been built similarly [63]. Another recent application of homology modeling is 3D
structure determination of RNA polymerase of the Ebola virus that helps in the detection of potential
therapeutic agents [64]. Furthermore 3D structure of NS5 protein of the Zika virus has been
determined by homology modeling that leads to the discovery of its potential inhibitors [65]. Recent
case applications of homology modeling in drug discovery are summarized in Table 2.

Opportunities and Possible Challenges in Homology Modeling

The number of high quality protein 3D structures has increased in the last decades. The introduction
of new experimental methods like Cryo-electron microscopy (Cryo-EM) is anticipated to increase the
number of 3D structures determined experimentally [80]. As the experimentally determined number
of high quality 3D protein structures of protein families increases, the role of homology modeling in
determining the 3D structures of the rest of the sequences in these families increases. However, 3D
structures of all protein distinct folds in nature has not been completed yet. As a result, there are
some difficulties in building 3D structures of proteins in which the structures of their protein families
have not been determined [18].

There are dozens of methods used for model building in homology modeling. New methods with new
algorithms have been developed. Various studies have demonstrated that there is no single modeling
program or server which is superior in every properties to others [81]. So, selecting the method/s to
be used according to the protein in hand and specific aim of future applications of the model is
important.

In classical homology modeling the model is built mainly based on sequence similarity. In the
experimental structure determination, ligands are absent as they are often lost during the
purification process. Thus, the resulting models that are built without considering the ligand

This article is protected by copyright. All rights reserved.

 information in the template represent an unliganded state. This shortcoming has been dealed with
the introduction of ligand sensitive approaches. However, such approaches need expertise and
Accepted Article manual interventions that takes time. Hence the introduction of fully automated homology modeling
tools that can deal with such problems is an important issue [82]. Furthermore, there are efforts to
integrate it with post modeling applications. For instance, there are works to integrate modeling
tools with thermostabilizing mutations [83].

Homology modeling may leave some unresolved questions in the computational models. This can be
reduced by using models that came from more experimentally determined structures which allow
better conceivable templates for targets. As consistent, accurate and progressive methods for the
improvement of models by shifting the coordinates parallel to the native state are developed,
coverage increases [84].

Another limitation of homology modeling is presence of loops and inserts as it is difficult to model
them without template data [85]. In order to have a model with high accuracy, optimization of the
loop region and side chains is important. Optimization encompasses refinement of the generated
models with molecular dynamics simulations. In case there is low sequence similarity level between
target and template, using multiple templates is advantageous. But using multiple templates may
lead to aberrations in the alignment unless templates which are from identical or analogous
phylogenetic tree are used as the target sequence [86]. Using PSI BLAST algorithm instead of normal
BLAST may provide optimal template selections in evolutionary distant cases.

At the end of the homology modeling process, many models of a target are built in general. Having
many models is an opportunity but identification of the best model needs further investigation. In
order to identify the best model, the constructed models are compared using various parameters.
Discrete Optimized Protein Energy (DOPE) score [87], Template Modeling (TM) score [88] and Root
Mean Square Deviation (RMSD) value [89] are used for comparison. The determinant parameter is
decided depending on the purpose of modeling results.

Conclusion

The gap between protein sequences available and protein 3D structures determined experimentally
is growing. Homology modeling aims at building 3D structure of proteins from their sequences by
using templates with an accuracy which is similar to the experimental methods. Thus, it has a big role
in filling the widening gap.

In recent years there are many advances in the tools and servers of homology modeling that improve
the accuracy of modeling results. This has an impact on each step of homology modeling. Better
alignment methods, loop modeling, side chain modeling and validation techniques have been
introduced. As the accuracy of models generated increases, their applications in the drug discovery
process increase. So, homology modeling contributes much in the drug discovery. Furthermore, in
the near future integration of homology modeling with other computer aided drug design methods
and post modeling applications are expected.

Homology modeling is used in determining 3D structures of proteins and it has many applications in
the drug discovery process.

This article is protected by copyright. All rights reserved.

 Homology URL address Short Discription
Modeling Tools
Accepted Article or Servers

MODELLER http://www.salilab.org/modelle Is a homology modeling tool that generates

r/ protein 3D structures with spatial restraints
method. It is available freely, has powerful
features and gives reliable results [33].

I-TASSER https://zhanglab.ccmb.med.umi Is a server that provides an internet based

ch.edu/I-TASSER/ service for protein structure prediction. It was
found to be one of the best methods in the
servers section of CASP experiments [26].

SWISS-MODEL http://swissmodel.expasy.org/ Is a server that gives protein 3D structure from

its amino acid sequence. It provides user friendly
web interface. This server uses model quality
estimation to select the most appropriate
templates and gives the expected accuracy of the
models built approximately [34].

Molecular https://www.chemcomp.com/ Is a combination of segment matching and

Operating MOE- modeling of insertion or deletion regions
Environment Molecular_Operating_Environm approaches. In addition to 3D structure
(MOE) ent.htm prediction, it has advanced loop modeling,
advanced alignment methods and powerful
alignment visualizer and editor [35].

Phyre2 http://www.sbg.bio.ic.ac.uk/ph This modeling uses various detection tools to

yre2/html/page.cgi?id=index generate 3D structures. It has special features
like ligand binding prediction and variant analysis
among the protein amino acid sequence [36].

HHpred http://toolkit.tuebingen.mpg.de This tool builds 3D structures using pairwise

/hhpred comparison of profile hidden Markov models
(HMMs) from a single or multiple query
sequence [37].

Robetta http://www.robetta.org/ Based on the ROSETTA fragment insertion

method, it gives both ab initio and homolog
models of protein regions [38].

Protein Model http://www.proteinmodelportal PMP provides interactive interface for model

Portal (PMP) .org/ building and quality assessment [39].

ICM https://www.molsoft.com/hom Is one of the homology modeling tools that give

ology.html 3D structure with good accuracy. Its features
include fast model building, loop prediction,
model validation and refinement [40].

Prime https://www.schrodinger.com/ Is a powerful package for accurate protein

prime structure prediction. In addition to building
structures with high accuracy, it provides
advanced simulation. It makes homology
modeling and fold recognition merge into a

This article is protected by copyright. All rights reserved.

 package. It has an easy –to-use interface [41].

SCWRL4 http://dunbrack.fccc.edu/scwrl4 It is a tool that is rapid with good accuracy and

Accepted Article
/index.php easy-to-use [42].

IntFOLD http://www.reading.ac.uk/bioin It is an independent server that predicts intrinsic

f/IntFOLD/ disorders, domains and protein-ligand binding
sites [43].

Protein (Molecule) Application Program/Server

Human angiotensin II type I Guide for designing novel BLAST, CLUSTALW, SYBYL,
receptor [66] therapeutic agents as angiotensin MODELLER, I-TASSER, PROCHECK,
receptor antagonists SurflexDock

HSP70 from GWD [67] Determination of 3D structure of BLAST, SWISS-MODEL, QMEAN,

hsp70 chaperone protein which is PSVS
a target of new wide spectrum
cancer therapeutics candidates

DNA Dependent Protein Kinase Screening of potential candidates WHATIF, ProSA, AutoDock Vina,
(DNA-PK) [68] as DNA-PK inhibitors Discovery Studio, ClustalW,
Phyre2, WHATCHECK

Human Concentrative Nucleoside Identification of sodium binding MOE, Gold, Glide, CHARMM
Transporter 3 (hCNT3) [69] site and the determinant residues
of nucleoside selectivity

Peroxisome Proliferator Identification of new ligand Prime, GlideXP, Schrodinger

Activated Receptor gamma molecules that reduce PPARγ
(PPARγ) [70] receptor in Type II diabetes
complications

GABA Transporter 1 (GAT1) [71] Discovering GAT1 inhibitor ClustalW, Prime, GlideXP,
molecules that are potential Schrodinger
anticonvulsant and
antidepressant agents

α-Glucosidase [72] Designing of new classes of α- BLAST, Prime, PROCHECK,

glucosidase inhibitors Sitemap, GlideXP, Schrodinger,
Maesterero

CD20 antigen [73] Insight into the structure of CD20 PSI-BLAST, T-Coffee, SWISS-
antigen which is a target in MODEL, I-TASSER, Phyre2,
developing new monoclonal MUSTER, Rampage
antibodies

Histamine H2 receptor [74] New perspectives into the BLAST, ClustalX, MODELLER,
development of a new potent PROCHECK, AutoDock, STRING
drugs against peptic ulcer by
targeting H2

Protein Kinase D 1 (PKD1) [75] Designing new PKD1 inhibitors Schrodinger suit, Maestero,
MODELLER, PDBSum, Glide XP

This article is protected by copyright. All rights reserved.

 Ribonucleotide reductase [76] Screening of novel drugs for drug SWISS-MODEL, HHPred, ProFunc,
resistant Leprosy therapy ERRAT, WHATIF, ProSA, Glide XP,
Accepted Article Schrodinger

Ecto-Nucleoside Triphosphate Structural insights into the binding MOE, BLAST, Rampage, Marvin
Diphosphohydrolases (E- of E-NTPDases to substrates and Sketch, FleX X, GROMACS
NTPDases) [77] inhibitors

Parkinson’s linked mutant Identification of multiple novel MOE, Glide 1, Maestero,

Leucine-Rich Repeat Kinase 2 points within neuronal death CHARMM
(LRRK2) [78] signaling pathways that could be
targeted by potential therapeutic
candidates

Alpha-Galactosidase A (α-Gal A) Detection of six GLA variants that Alamut Visual, SWISS MODEL,
[79] cause α-Gal A activity deficiency PyMol
and protein wild type structure
loss

This article is protected by copyright. All rights reserved.

Accepted Article

This article is protected by copyright. All rights reserved.