Development, evaluation and application of 3D QSAR Pharmacophore model in the discovery of potential human renin inhibitors.

Abstract

Renin has become an attractive target in controlling hypertension because of the high specificity towards its only substrate, angiotensinogen. The conversion of angiotensinogen to angiotensin I is the first and rate-limiting step of renin-angiotensin system and thus designing inhibitors to block this step is focused in this study.
Ligand-based quantitative pharmacophore modeling methodology was used in identifying the important molecular chemical features present in the set of already known active compounds and the missing features from the set of inactive compounds. A training set containing 18 compounds including active and inactive compounds with a substantial degree of diversity was used in developing the pharmacophore models. A test set containing 93 compounds, Fischer randomization, and leave-one-out methods were used in the validation of the pharmacophore model. Database screening was performed using the best pharmacophore model as a 3D structural query. Molecular docking and density functional theory calculations were used to select the hit compounds with strong molecular interactions and favorable electronic features.
The best quantitative pharmacophore model selected was made of one hydrophobic, one hydrogen bond donor, and two hydrogen bond acceptor features with high a correlation value of 0.944. Upon validation using an external test set of 93 compounds, Fischer randomization, and leave-one-out methods, this model was used in database screening to identify chemical compounds containing the identified pharmacophoric features. Molecular docking and density functional theory studies have confirmed that the identified hits possess the essential binding characteristics and electronic properties of potent inhibitors.
A quantitative pharmacophore model of predictive ability was developed with essential molecular features of a potent renin inhibitor. Using this pharmacophore model, two potential inhibitory leads were identified to be used in designing novel and future renin inhibitors as antihypertensive drugs.

Full text

Page 1

PROCEEDINGSOpen Access
Development, evaluation and application of 3D
QSAR Pharmacophore model in the discovery of
potential human renin inhibitors
Shalini John, Sundarapandian Thangapandian, Mahreen Arooj, Jong Chan Hong, Kwang Dong Kim,
Keun Woo Lee*
From 22nd International Conference on Genome Informatics
Busan, Korea. 5-7 December 2011
Abstract
Background: Renin has become an attractive target in controlling hypertension because of the high specificity
towards its only substrate, angiotensinogen. The conversion of angiotensinogen to angiotensin I is the first and
rate-limiting step of renin-angiotensin system and thus designing inhibitors to block this step is focused in this
study.
Methods: Ligand-based quantitative pharmacophore modeling methodology was used in identifying the
important molecular chemical features present in the set of already known active compounds and the missing
features from the set of inactive compounds. A training set containing 18 compounds including active and inactive
compounds with a substantial degree of diversity was used in developing the pharmacophore models. A test set
containing 93 compounds, Fischer randomization, and leave-one-out methods were used in the validation of the
pharmacophore model. Database screening was performed using the best pharmacophore model as a 3D
structural query. Molecular docking and density functional theory calculations were used to select the hit
compounds with strong molecular interactions and favorable electronic features.
Results: The best quantitative pharmacophore model selected was made of one hydrophobic, one hydrogen
bond donor, and two hydrogen bond acceptor features with high a correlation value of 0.944. Upon validation
using an external test set of 93 compounds, Fischer randomization, and leave-one-out methods, this model was
used in database screening to identify chemical compounds containing the identified pharmacophoric features.
Molecular docking and density functional theory studies have confirmed that the identified hits possess the
essential binding characteristics and electronic properties of potent inhibitors.
Conclusion: A quantitative pharmacophore model of predictive ability was developed with essential molecular
features of a potent renin inhibitor. Using this pharmacophore model, two potential inhibitory leads were identified
to be used in designing novel and future renin inhibitors as antihypertensive drugs.
* Correspondence: kwlee@gnu.ac.kr
Division of Applied Life Science (BK21 Program), Systems and Synthetic
Agrobiotech Center (SSAC), Research Institute of Natural Science(RINS), Plant
Molecular Biology and Biotechnology Research Center (PMBBRC),
Gyeongsang National University (GNU), 501 Jinju-daero, Gazha-dong, Jinju
660-701, Republic of Korea
John et al. BMC Bioinformatics 2011, 12(Suppl 14):S4
http://www.biomedcentral.com/1471-2105/12/S14/S4
© 2011 John et al; licensee BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons
Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in
any medium, provided the original work is properly cited.

Page 2

Background
Hypertension is a major factor concerning various cardi-
ovascular diseases such as congestive cardiac failure,
stroke, and myocardial infarction and affects up to 30%
of the adult population in most countries [1]. Renin is an
aspartyl protease and catalytically similar to other
enzymes such as pepsin, cathepsin and chymosin etc [2].
Renin cleaves the angiotensinogen to angiotensin-I which
is then converted to angiotensin-II by the action of
angiotensinogen converting enzyme (ACE). Angiotensin-
II is a biologically active vasopressor recognized by its
receptors which is one of the cascades of events that
leads to the increase in blood pressure. Renin is synthe-
sized as prorenin, a proenzyme, which is transformed
into mature renin by the cleavage of 43 amino acids long
prosegment from the N-terminal end. This conversion of
prorenin to renin occurs in the juxtaglomerular cells of
kidney followed by the release of renin into the circula-
tion [3]. Renin blocks the first and rate-limiting step
which is the conversion of angiotensinogen to angioten-
sin-I. Renin is a very specific enzyme towards its only
known substrate, angiotensinogen, and this remarkable
specificity makes it a very attractive and ideal target to
block the renin-angiotensin system (RAS) [4]. Inhibition
of renin prevents the formation of both angiotensin-I and
II but this is not the case in ACE inhibitors and angioten-
sin receptor blockers, which increase angiotensin-I or/
and II level, respectively. Only renin inhibitors will render
the complete RAS quiescent by suppressing the first step
of the cascade of events. Thus, inhibition of renin would
favor more complete blockade of the system [5]. Potent
inhibitors of this enzyme could therefore provide a new
alternative way to treat hypertension without inhibiting
other biological substances. Aspartyl protease class of
enzymes contains two aspartic acid residues that are
necessary for the activity. Renin enzyme has a bilobal
structure similar to other aspartic proteases and an active
site at the interface. The two important aspartate residues
Asp32 and Asp215 catalyze the proteolytic function of
renin are donated from each lobes of the enzyme [6].
The active site of renin appears as a long, deep cleft that
can accommodate seven amino acid units of the sub-
strate, angiotensinogen, and cleaves the peptide bond
between Leu10 and Val11 within angiotensinogen to gen-
erate angiotensin-I [7]. The approaches followed to
develop early renin inhibitors were based on two meth-
odologies. One is to develop similar peptides to prorenin
as this segment covers the active site of renin prior to the
maturation. The second is based on the N-terminal por-
tion of the substrate, angiotensinogen, for this binds the
active site of renin. But these approaches produced only
weak inhibitors [8]. The first synthetic renin inhibitor
was pepstatin. First-generation renin inhibitors were
peptide analogues of the prosegment of renin or sub-
strate analogues of the amino-terminal sequence of
angiotensinogen containing the renin cleavage site [9].
Crystal structure analyses of renin-inhibitor complexes
and computational molecular modeling were later used
to design selective nonpeptide renin inhibitors that
lacked the extended peptide-like backbone of previous
inhibitor sand had improved pharmacokinetic properties
[10]. Aliskiren is the first of these new nonpeptide inhibi-
tors to be approved by the FDA for the treatment of
hypertension but its synthesis include many steps. This
invites much simpler compounds to be designed as
potent renin inhibitors [11]. Aliskiren belongs to the
third generation of renin inhibitors where the large (high
molecular weight) first and second generation inhibitors
could not be exploited as drugs despite of their potency
in vitro [12]. To date, only few compounds were success-
fully developed with potent renin inhibition profiles, high
efficacy, and safety. Thus designing inhibitors of high
potential for renin inhibition is the most effective way to
block the RAS completely. This study was focused to
identify novel scaffolds with the potential to turn as the
new category of renin inhibitors.
A high-correlation quantitative pharmacophore model
was generated, in this study, using the observed struc-
ture-activity relationship of known renin inhibitors. We
have successfully applied pharmacophore modeling, data-
base screening, molecular docking, and density functional
theory (DFT) calculation methodologies in identifying
lead candidates to be employed in potent renin inhibitor
design and thereby new category of anti-hypertensive
agents.
Methods
Selection of training set compounds
Three dimensional (3D) QSAR strategy is one of the
ligand-based pharmacophore modeling approaches. This
strategy differs from the common feature pharmaco-
phore approach in various points such as limitations of
number of training set compounds and necessity of
experimental activity values predicted using similar
bioassay conditions etc [13]. More than 300 chemical
compounds were retrieved from various literature
resources [14-19] and 111 compounds evaluated with
the same bio-assay protocol were selected to be used as
primary data set in 3D QSAR pharmacophore modeling
study. To ensure the statistical relevance, a training set
containing 18 diverse compounds with the experimental
activity values (IC50) ranging from 0.5 nM to 5590 nM
were selected from 111 dataset compounds and used as
training set (Figure 1) and the remaining 93 compounds
were used as test set compounds to be utilized in phar-
macophore validation.
John et al. BMC Bioinformatics 2011, 12(Suppl 14):S4
http://www.biomedcentral.com/1471-2105/12/S14/S4
Page 2 of 14

Page 3

Compounds preparation and conformation generation
The two-dimensional (2D) chemical structures of all the
compounds in the data set were sketched using ChemS-
ketch, version 12 (ACD Inc., Toronto, Canada) and subse-
quently converted to 3D structures in Accelrys Discovery
Studio 2.5 (DS). These 3D compounds were further
checked for the added hydrogens and minimized using
smart minimizer that performs 1000 steps of steepest des-
cent followed by conjugate gradient algorithms with a con-
vergence gradient of 0.001 kcal mol-1. After energy
minimization, multiple acceptable conformers were gener-
ated for every training set compound within DS Diverse
Conformation Generation module using the Poling algo-
rithm. This step was necessary to produce a good set of
representative conformations of different conformation
space accessible to a molecule within a given energy range.
A maximum of 255 conformations were generated for
each compound within an energy range of 20 kcal mol-1
above the global energy minimum [20-22].
Generation of pharmacophore models
Among the two types of ligand-based pharmacophore
modeling methodologies, common feature pharmaco-
phore modeling utilizes the common chemical features
present only in the most active compounds whereas the
3D QSAR pharmacophore methodology uses the chemi-
cal features of most active and inactive compounds along
with their biological activity. In this study, we have
employed 3D QSAR-based pharmacophore methodology
to generate pharmacophore models that can be used to
Figure 1 Structure of the training set compounds. 2D Chemical structures of the 18 training set compounds together with their
experimental IC50values.
John et al. BMC Bioinformatics 2011, 12(Suppl 14):S4
http://www.biomedcentral.com/1471-2105/12/S14/S4
Page 3 of 14

Page 4

estimate the activity of newly designed compounds.
Feature mapping protocol as available in DS was used to
identify the features that are present in the training set
compounds. Uncertainty value was set to 2 and the mini-
mum inter-feature distance was set to 2Å from the
default value of 2.97 Å. As identified by the feature map-
ping protocol, hydrogen bond acceptor (HBA), hydrogen
bond donor (HBD), hydrophobic aliphatic (HY-AL),
hydrophobic aromatic (HY-AR) and ring aromatic (RA)
features were used with other default values to generate
ten pharmacophore models using 3D QSAR pharmaco-
phore generation of DS. Each feature of the resulting
models occupies a certain weight that is proportional to
its relative contribution to biological activity. HypoGen
therefore constructs pharmacophore models correlating
best with biological activities and consisting of as few fea-
tures as possible. The HypoGen pharmacophore model
generation process is performed in three steps such as
the constructive phase, the subtractive phase and the
optimization phase [23,24]. Hypotheses that are common
to the most active set of compounds are identified during
the constructive phase. HypoGen calculates all possible
pharmacophore configurations using all combinations of
pharmacophore features for each of the conformations of
the two most active compounds. Additionally, the
hypotheses must fit a minimum subset of features of the
remaining most active compounds in order to be consid-
ered. A large database of pharmacophore configurations
is generated at the end of the constructive phase. In the
subtractive phase, all pharmacophore configurations that
are also present in the least active set of molecules are
removed. All compounds whose activity is by default 3.5
orders of magnitude less than that of the most active
compound are considered to represent the least active
molecules. The value 3.5 is adjustable depending on the
activity range of the training set. During the optimization
phase, the hypothesis score is improved. Hypotheses are
scored based on errors in activity estimates from regres-
sion and complexity. The optimization involves a variation
of features and/or locations to optimize activity prediction
via a simulated annealing approach. When the optimiza-
tion process no longer improves the score, HypoGen stops
and reports the top scoring 10 unique pharmacophores.
The generated pharmacophore models were evaluated for
their reliability based on the cost parameters. The overall
costs of a model consist of three cost components, namely,
the weight cost, the error cost, and the configuration cost.
The weight component is a value that increases in a Gaus-
sian form as this function weights in a model deviate from
the ideal value of two. The error cost represents the differ-
ence between estimated and measured activities of the
training set. The configuration cost quantifies the entropy
of the hypothesis space.
In addition, the following three cost values are calculated
during the generation of pharmacophore models: the fixed
cost, the total cost, and the null cost. The fixed cost is the
lowest possible cost representing a hypothetical simplest
model that fits all data perfectly. Fixed costs are calculated
by adding the minimum achievable error and weight cost
and the constant configuration cost. Another cost para-
meter, the null cost, represents the maximum cost of a
pharmacophore with no features and estimates activity to
be the average of activity data of training set molecules.
The null cost value is equal to the maximum occurring
error cost. For every pharmacophore generation ten total
cost values and each of fixed cost and null cost values are
calculated by the pharmacophore generation protocol in
the unit of bits. For a meaningful pharmacophore model,
the fixed cost should be lower and the null cost should be
higher and the total cost value should be closer to the
fixed cost and away from the null cost value [25,26].
HypoGen further estimates the activity of each training set
compound using regression parameters. The parameters
were computed by regression analysis using the relation-
ship of geometric fit value versus the negative logarithm of
activity. The better the geometric fit the greater the activ-
ity prediction of the compound. Along with these cost
values, other statistical values such as correlation coeffi-
cient and root mean square deviation (RMSD) were calcu-
lated. The best pharmacophore model was selected based
on the large cost difference, high correlation coefficient
and lower RMSD.
Pharmacophore validation
The main purpose to validate the generated pharmaco-
phore models is to investigate their ability to estimate the
activity of new compounds identified through database
screening or designed de novo. The selected pharmaco-
phore model was validated using three methods based on
the derived cost components, ability in test set prediction,
Fischer randomization test results, and leave-one-out
method. A larger difference between the fixed and null
costs than that between the fixed and total costs signifies
the quality of a pharmacophore model. All of these cost
values are reported in bits and a difference of 40-60 bits
between the total and null costs suggests a 75-90% chance
of representing a true correlation in the data [27,28].
Ninety three diverse compounds were used as the test set
to validate the pharmacophore model. Fischer randomiza-
tion is another approach for pharmacophore model valida-
tion. The 95% confidence level was selected in this
validation study and 19 random spreadsheets were con-
structed. This validation method checks the correlation
between the chemical structures and biological activity.
This method generates pharmacophore models using the
same parameters as those used to develop the original
John et al. BMC Bioinformatics 2011, 12(Suppl 14):S4
http://www.biomedcentral.com/1471-2105/12/S14/S4
Page 4 of 14

Page 5

pharmacophore model by randomizing the activity data of
the training set compounds. Finally the cross validation of
the model was performed by using the leave-one-out
methodology. In this method, 18 pharmacophore models
were generated with the same parameters used for gener-
ating original pharmacophore model but leaving one com-
pound at a time from 18 training set compounds to
ensure the influence of every single training set compound
in the generation of selected pharmacophore model
[29,30].
Database screening and drug-like prediction
The best pharmacophore model validated using different
methods was used as a 3D query in database screening to
retrieve chemical compounds that fit all the pharmaco-
phoric features. A chemical compound must fit all the fea-
tures to be picked as hits. Search 3D Database protocol
with Best/Flexible search option was employed in database
screening. Three chemical databases of diverse chemical
compounds were screened for novel chemical scaffolds to
be used in potent renin inhibitor design. The identified
database hits were screened using various filters based on
estimated activity, Lipinski’s rule of five [31], and ADMET
properties [32-35].
Molecular docking
Compounds satisfying all the filters were subjected to
molecular docking studies. The GOLD (Genetic Optimiza-
tion for Ligand Docking) program from Cambridge Crys-
tallographic Data Centre, UK uses a genetic algorithm to
dock the small molecules into the protein active site was
used in molecular docking [36]. GOLD allows for a full
range of flexibility for the ligands and partial flexibility of
the protein. Protein coordinates from the crystal structure
complex of renin with aliskiren (PDB ID: 2V0Z), one of
the most active inhibitors, determined at a resolution of
2.20 Å were used to define the active site. The active site
was defined with a 10 Å radius around the bound inhibi-
tor. All the water molecules except two catalytically
important 184 and 250 were removed from the protein
and hydrogens were added. The ten top-scoring confor-
mations of every ligand were saved at the end of the calcu-
lation. Early termination option was used to skip the
genetic optimization calculation when any five conforma-
tions of a particular compound were predicted within an
RMSD value of 1.5 Å. The GOLD fitness score is calcu-
lated from the contributions of hydrogen bond and van
der Waals interactions between the protein and ligand,
intramolecular hydrogen bonds and strains of the ligand
[37,38]. Protein-ligand interactions were analyzed using
DS and Molegro Virtual Docker [39] programs. The
novelty of the final hits was confirmed using SciFinder
[40] and PubChem [41] structure search tools.
Density functional theory (DFT) calculations
The final hits along with some most and least active
compounds were used as input and all DFT calculations
were carried out using Gaussian version 3.0 program.
The geometry optimization of a set of compounds was
carried out using the Becke3 Lee-Yang-Parr correlation
functional (B3LYP), at the 6-31G* level [42-45]. The orbi-
tal energies of frontier orbitals, namely, highest occupied
molecular orbital (HOMO) and lowest unoccupied mole-
cular orbital (LUMO) were calculated for a set of com-
pounds. The calculation was performed to evaluate the
electronic properties of final hits to be compared with
the compounds in the training set [46].
Results and discussion
Pharmacophore generation
A set of ten pharmacophore models was generated using
a training set containing 18 compounds by selecting
HBA, HBD, HY-AL, HY-AR and RA features as sug-
gested by Feature Mapping protocol. All the generated
pharmacophore models composed of either HBA or
HBD or both with HY-AL or HY-AR features. The total
cost values of ten pharmacophore models ranged from
81.50 to 99.54. The cost difference between the total
cost and null cost must be greater and it should be
smaller between total cost and fixed cost values for a
significant pharmacophore model. In our study, the
pharmacophore generation run calculated a fixed cost
value of 70.08 and the null cost value of 148.56. Among
the total cost values of generated ten pharmacophore
models, first model (Hypo1) has scored the value closer
to the fixed cost value when compared to other models.
The cost difference between the null cost and total cost
value of the first pharmacophore model is 67.06 (Table
1). The cost difference value between 40 and 60 implies
that the pharmacophore model correlates the experi-
mental and estimated activity values more than 90%. In
this study, the cost difference value of Hypo1 signifies
that it can correlate the experimental and estimated
activity values of the training set compounds more than
90%. Hypo1 was made of four pharmacophoric features
consisting two HBA, one HBD and one HY-AL features
(Figure 2).
Further evaluation of the generated pharmacophore
models was based on the correlation coefficient. The
correlation values of these ten pharmacophore models
were greater than 0.840, and the first three pharmaco-
phore models correlated the activity data with high cor-
relation values, i.e., above 0.9. These results indicate the
capability of the pharmacophore model to predict the
activity of the training set compounds. Hypo1 showed
the highest correlation coefficient value of 0.944, high-
lighting its strong predictive ability. In addition, RMSD
John et al. BMC Bioinformatics 2011, 12(Suppl 14):S4
http://www.biomedcentral.com/1471-2105/12/S14/S4
Page 5 of 14