Ligand efficiency based approach for efficient virtual screening of compound libraries

Article abstract of DOI:10.1016/j.ejmech.2014.06.029

Here we report for the first time the use of fit quality (FQ), a ligand efficiency (LE) based measure for virtual screening (VS) of compound libraries. The LE based VS protocol was used to screen an in-house database of 125,000 compounds to identify aurora kinase A inhibitors. First, 20 known aurora kinase inhibitors were docked to aurora kinase A crystal structure (PDB ID: 2W1C); and the conformations of docked ligand were used to create a pharmacophore (PH) model. The PH model was used to screen the database compounds, and rank (PH rank) them based on the predicted IC₅₀ values. Next, LE-Scale, a weight-dependant LE function, was derived from 294 known aurora kinase inhibitors. Using the fit quality (FQ = LE/LE-Scale) score derived from the LE-Scale function, the database compounds were reranked (PH-FQ rank) and the top 151 (0.12% of database) compounds were assessed for aurora kinase A inhibition biochemically. This VS protocol led to the identification of 7 novel hits, with compound 5 showing aurora kinase A IC₅₀ = 1.29 μM. Furthermore, testing of 5 against a panel of 31 kinase reveals that it is selective toward aurora kinase A & B, with <50% inhibition for other kinases at 10 μM concentrations and is a suitable candidate for further development. Incorporation of FQ score in the VS protocol not only helped identify a novel aurora kinase inhibitor, 5, but also increased the hit rate of the VS protocol by improving the enrichment factor (EF) for FQ based screening (EF = 828), compared to PH based screening (EF = 237) alone. The LE based VS protocol disclosed here could be applied to other targets for hit identification in an efficient manner.

Full text of DOI:10.1016/j.ejmech.2014.06.029

Accepted Manuscript
Ligand efficiency based approach for efficient virtual screening of compound libraries
Yi-Yu Ke, Mohane Selvaraj Coumar, Hui-Yi Shiao, Wen-Chieh Wang, Chieh-Wen
Chen, Jen-Shin Song, Chun-Hwa Chen, Wen-Hsing Lin, Szu-Huei Wu, John T. A
Hsu, Chung-Ming Chang, Hsing-Pang Hsieh
PII:
S0223-5234(14)00548-0
10.1016/j.ejmech.2014.06.029
EJMECH 7072
DOI:
Reference:
To appear in: European Journal of Medicinal Chemistry
Received Date: 25 February 2014
Revised Date: 4 June 2014
Accepted Date: 13 June 2014
Please cite this article as: Y.-Y. Ke, M.S. Coumar, H.-Y. Shiao, W.-C. Wang, C.-W. Chen, J.-S. Song,
C.-H. Chen, W.-H. Lin, S.-H. Wu, J.T.A. Hsu, C.-M. Chang, H.-P. Hsieh, Ligand efficiency based
approach for efficient virtual screening of compound libraries, European Journal of Medicinal Chemistry
(2014), doi: 10.1016/j.ejmech.2014.06.029.
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT
Graphical abstract
Application of Fit quality (FQ) - a ligand efficiency based measure for virtual screening
had improved the rate of hit identification from the database, compared to pharmacophore
based screening alone.
1

ACCEPTED MANUSCRIPT
Highlights
•
•
•
•
Fit quality (FQ) - a ligand efficiency based measure for virtual screening is described
FQ assisted screening of a database for aurora kinase inhibitor undertaken
The protocol identified compound 5 as selective aurora kinase inhibitor
The application of FQ for virtual screening improved the hit rates significantly

ACCEPTED MANUSCRIPT
Submitted to EJMC as an Original Research Article
Ligand efficiency based approach for efficient virtual screening of compound libraries
a,*
b
a
a
a
Yi-Yu Ke , Mohane Selvaraj Coumar , Hui-Yi Shiao , Wen-Chieh Wang , Chieh-Wen Chen ,
a
a
a
a
a
Jen-Shin Song , Chun-Hwa Chen , Wen-Hsing Lin , Szu-Huei Wu , John T. A. Hsu , Chung-
a
a,∗
Ming Chang , Hsing-Pang Hsieh
a
Institute of Biotechnology and Pharmaceutical Research, National Health Research
Institutes, 35 Keyan Road, Zhunan, Miaoli County 350, Taiwan, ROC.
b
Centre for Bioinformatics, School of Life Sciences, Pondicherry University, Kalapet,
Puducherry 605014, India
∗
Corresponding author: phone: +886-37-246-166 ext. 35708; fax: +886-37-586-456; e-mail:
yiyuke@nhri.org.tw, hphsieh@nhri.org.tw.
1

ACCEPTED MANUSCRIPT
Abstract
Here we report for the first time the use of fit quality (FQ), a ligand efficiency (LE) based
measure for virtual screening (VS) of compound libraries. The LE based VS protocol was
used to screen an in-house database of 125,000 compounds to identify aurora kinase A
inhibitors. First, 20 known aurora kinase inhibitors were docked to aurora kinase A crystal
structure (PDB ID: 2W1C); and the conformations of docked ligand were used to create a
pharmacophore (PH) model. The PH model was used to screen the database compounds, and
rank (PH rank) them based on the predicted IC50 values. Next, LE_Scale, a weight-dependant
LE function, was derived from 294 known aurora kinase inhibitors. Using the fit quality (FQ
=
LE/LE_Scale) score derived from the LE_Scale function, the database compounds were
reranked (PH_FQ rank) and the top 151 (0.12% of database) compounds were assessed for
aurora kinase A inhibition biochemically. This VS protocol led to the identification of 7 novel
hits, with compound 5 showing aurora kinase A IC50 = 1.29 ꢀM. Furthermore, testing of 5
against a panel of 31 kinase reveals that it is selective toward aurora kinase A & B, with <
50% inhibition for other kinases at 10 ꢀM concentrations and is a suitable candidate for
further development. Incorporation of FQ score in the VS protocol not only helped identify a
novel aurora kinase inhibitor, 5, but also increased the hit rate of the VS protocol by
improving the enrichment factor (EF) for FQ based screening (EF = 828), compared to PH
based screening (EF = 237) alone. The LE based VS protocol disclosed here could be applied
2

ACCEPTED MANUSCRIPT
to other targets for hit identification in an efficient manner.
Keywords: Ligand efficiency, Fit quality, Virtual screening, Pharmacophore model, Aurora
kinase inhibitor
3

ACCEPTED MANUSCRIPT
1. Introduction
Using the advancements in parallel synthesis technology of the late 1980s and early 1990s,
several pharmaceutical companies set up high-throughput screening (HTS) of large
compound libraries in order to identify leads for drug development. One of the main
disadvantages of HTS is the need to screen a large collection of compounds, which is both
time and resource intensive. Recently, virtual screening (VS) is routinely used as a
supplement to or as a replacement for HTS in lead identification, due to virtual screening’s
savings in time and money. VS is a computational method used to score and prioritize HTS
library compounds for further biochemical testing [1, 2]. Use of VS can increase the hit rates
of lead identification, with some studies reporting hit rates ten times greater or more for VS
compared to HTS [3].
Several success stories of lead identification using VS are reported in the literature [3,
4], using either structure-based virtual screening (SBVS) [5] or ligand-based virtual screening
(LBVS) [6]. In SBVS, molecular docking of the compounds with the target protein is
performed and a score is assigned based on the interaction of the target protein and the
compounds; the score of the compounds is used to rank and prioritize them for testing.
However, current docking methods have deficiencies in both correctly identifying the
crystallographic conformation of the ligands and also accurately scoring the docked poses [7-
9]. More critically, availability of target protein information (either 3D X-ray
4

ACCEPTED MANUSCRIPT
crystallographic/NMR solved structure or amino acid sequence of the target protein) is
essential for SBVS. Hence, in the absence of such protein information, LBVS (an alternative
method for VS) is used to identify leads for drug-development projects.
In LBVS, ligand similarities based on either 2D or 3D information of known ligands
are used to identify new leads [6]. Particularly, 3D-QSAR (quantitative structure-activity
relationship) based techniques — such as CoMFA (comparative molecular field analysis) [10],
[
[
11], CoMSIA (comparative molecular similarity indices) [12], and pharmacophore models
13] — are currently much used. A pharmacophore model represents the relative orientation
of the functional groups of a molecule in three-dimensional space, which are necessary for
maintaining the activity. A typical pharmacophore model constitutes various pharmacophoric
elements such as the hydrogen bond acceptor (HBA), hydrogen bond donor (HBD),
hydrophobic (HY) group, and excluded volume (EV); the relative positions of the
pharmacophoric elements in the model are defined by the distances and angles between them.
Pharmacophore models can be generated from a set of known active ligands, by
superimposition of the bioactive conformation of the ligands. Recently, several groups have
reported the use of pharmacophore models for carrying out VS to identify hits [13].
Hits identified from VS, either from SBVS or LBVS methods, are used as a starting
point for lead optimization in drug discovery programs. Retrospective analysis of the starting
lead ‒ drug pairs from several successful drug discovery programs reveals that the molecular
5

ACCEPTED MANUSCRIPT
properties (molecular weight and lipophilicity) inflation is very common during lead
optimization; hence there is a need to keep tight control over them during the process of drug
evolution [14]. Moreover, the success of developing a drug depends heavily on the nature of
the starting hits chosen for lead optimization. Simple criteria such as activity alone are
insufficient for selecting an appropriate hit for optimization; additional criteria such as
Ligand efficiency (LE) need to be considered to select the right hit candidates for further
optimization [15, 16]. LE can be defined as the binding affinity per heavy atom of the
molecule and can be computed from experimental or calculated binding affinity divided by
the number of heavy atoms. LE was first introduced in the field of fragment-based drug
discovery (FBDD) to select the best of two fragments and is now routinely used to choose the
optimized leads during hit to drug evolution in drug discovery programs [17]. It should also
be noted that the application of LE has its own limitation as it considers all heavy atoms the
same even though oxygen, nitrogen, and other important heavy atoms present in drugs have
different physico-chemical and binding properties. This limitation of LE is recognized in the
field of drug design; accordingly, the use of the LE function in conjunction with other
parameters such as LLE (lipophilic ligand efficiency) during hit selection/optimization in
drug discovery programs is recommended [18-20].
The aim of this work is to develop an efficient VS protocol for the improved
identification of hits from an HTS library; the VS protocol is exemplified by aurora kinase
6

ACCEPTED MANUSCRIPT
inhibitor identification. For this purpose, 125,000 compounds in an in-house HTS library
were prioritized for biochemical screening using a two-step procedure (Scheme 1). First, HTS
compounds were screened by a pharmacophore model developed using two sets of aurora
kinase inhibitors (pyrazoles & furanopyrimidines). The model predicts the IC50 of the HTS
compounds by matching them to the pharmacophore features of the model and ranks
(PH_rank) them based on their predicted activity. In the second step, compounds identified
using the pharmacophore model were reranked (PH_FQ rank) using Fit quality (FQ) score, a
LE based function. Finally, the top ranked compounds were subjected to biochemical testing
for aurora kinase A inhibition. The VS protocol reported here has led to the identification of a
potent aurora kinase A inhibitor with an IC50 = 1.29 ꢀM, by testing only 151 compounds from
a HTS library of 125,000 compounds. Moreover, incorporation of a LE based scoring
function into the VS method has improved the hit rate; such a protocol could be applied to
other targets as well.
2. Chemistry
The aurora kinase inhibitor 5 identified in this study was synthesized from commercially
available methyl-5-formyl-2-hydroxybenzoate (8) and p-nitro-aniline (14) using a convergent
synthesis protocol as shown in Scheme 2. For this purpose, aldehyde 8 was subjected to
bromination under acid condition, as reported earlier, to give 9 in quantitative yield [21].
Treatment of 9 with the Grignard reagent ‒ 2-chlorophenylmagnesium bromide in THF at 0°
7

ACCEPTED MANUSCRIPT
C provided the secondary alcohol 10 in 65% yield. Acid mediated dehydroxylation of 10
followed by LAH reduction resulted in the primary alcohol 12 in 83% yields, over two steps.
2
The key aldehyde intermediate 13 was prepared from 12 by MnO oxidation, in 77% yield.
Next, the aniline intermediate 16 was synthesized from the aniline 14, by first coupling with
-furoyl chloride under basic condition to give 15 in 74% yield, followed by hydrogenolysis
2
2
under H atmosphere over 10% Pd/C in 95% yield. Finally, both intermediates, 13 and 16,
were condensed under reflux conditions to give the desired compound 5 in 52% yield.
3. Results and Discussion
3.1. Pharmacophore model generation and validation
As a first step, we set out to develop an LBVS platform using pharmacophore model for
screening the in-house HTS library to identify aurora kinase inhibitors. It is known that 3D
pharmacophore based VS is much faster than SBVS. Moreover, pharmacophore-based VS
results in the identification of much more diverse chemotypes and is more useful in scaffold-
hopping [13]. Albeit these advantages of LBVS, use of SBVS is often much higher than the
use of LBVS (pharmacophore) methods for hit identification, due to certain limitations [4].
For instance, the accuracy of the pharmacophore model generated from a set of known active
ligands depends on the ability of the algorithm to correctly predict the bioactive conformation
of the ligands and then align them. Previous studies by several groups have shown that the
docking-based alignment of ligands can overcome this shortcoming and has resulted in better
8

ACCEPTED MANUSCRIPT
3D-QSAR models [22, 23]. Hence, in our studies we first planned to use structure-based
docking to generate the bioactive conformations of the ligands for alignment and
subsequently use them for pharmacophore model building.
The aurora kinase inhibitors used for developing the pharmacophore model were the
pyrazole and furanopyrimidine analogs reported by us earlier (Figure 1A) [24-27]. Selection
of training and test set compounds (Table S1 and S2, supplemental information) were
performed using the Discovery Studio/Cluster Ligands program, as described before for the
construction of COMFA and COMSIA 3D-QSAR models [27]. Pharmacophore hypothesis
generation was performed by the HypoGen module of Discovery Studio automatically, based
on the structure-based alignment derived from docked conformations of 20 training set
ligands to the aurora kinase A crystal structure (PDB ID: 2W1C). A total of 20 hypotheses
were generated by the HypoGen; the top 10 hypothesis with their statistical parameters and
features are shown in Table S3, supplemental information. The quality of pharmacophore
hypothesis can be evaluated using various cost functions: total cost, fixed cost, null cost,
configuration cost, and error cost [28]. Even though the top three hypothesis (Hypo 1, 2 and 3)
have higher correlation values, these three models do not have the important hydrophobic
feature corresponding to the back pocket region of the protein (Figure S1, supplemental
information). Hydrophobic interaction in the back pocket region of Aurora kinase was
identified as essential feature for maintaining potent Aurora kinase inhibition activity in our
9

ACCEPTED MANUSCRIPT
previous studies [24-27]. Hence, the hypothesis 4 (Hypo 4) was chosen as the best, with its
total cost of 80.00, fixed cost of 51.80, null cost of 165.72, error cost of 70.56, RMS of 1.71,
correlation of 0.88, and configuration cost of 7.19. This hypothesis has 90% statistical
significance, since the cost differences (ꢁ Cost) between null and fixed or null and total cost
were both greater than 60, implying that hypothesis 4 has a greater than a 90% chance of
representing a true correlation between the pharmacophoric features represented by the
hypothesis and the aurora kinase A activity. Moreover, the configuration cost of the top
hypothesis generated was smaller than 17, which indicates that the hypothesis was not
generated by chance and is unlikely to correlate with others.
To further confirm that the hypothesis generated was not by chance correlation,
Fischer’s randomization test with 95% confidence level was carried out by randomly
assigning the activities to the training set molecules. A total of 19 spreadsheets were
generated from the randomization of data, but none of them showed statistically significant
correlation with the activity data, further confirming that hypothesis 4 is a genuine model
(Table S4, supplemental information).
In our previous studies [24-26], both the pyrazole and furanopyrimidine aurora kinase
A inhibitors (training set) were found to have the following major structural features that are
essential for interaction at the binding site of the enzyme: 1) the N1 nitrogen of
furanopyrimidine or pyrazole ring makes hydrogen bond with the protein, 2) the urea side
10

ACCEPTED MANUSCRIPT
chain CO group of furanopyrimidine makes another hydrogen bond with the protein, and 3)
the urea side chain terminal phenyl group makes hydrophobic interaction in the back pocket.
These structural requirements for interaction with aurora kinase A match well with the
pharmacophoric features of hypothesis 4 generated by HypoGen (Figure 1B), which includes
one hydrogen bond acceptor (HBA), one hydrogen bond donor (HBD), and three
hydrophobic features (HY1-HY3).
Next, the prediction ability of this pharmacophore model (Hypothesis 4) was
evaluated by the goodness of prediction of aurora kinase A activity for a set of 46 test set
compounds (Test set - 1: pyrazoles and furanopyrimidines, Table S2, supplemental
information) and 9 aurora kinase inhibitors in clinical trials (Test set - 2: AT9283 - Phase I
[
29] , AZD-1152 - Phase I [30], GSK-1070916 - Phase I [31], MLN-8237 - Phase II [32], PF-
0
3814735 - Phase I [33], PHA-680632 – Preclinical [34], PHA-739358 - Phase II [35], SNS-
14 - Phase I [36], and ZM-447439 - Phase I [37], Table S5, supplemental information). The
3
linear regression coefficient for the activity prediction (r ꢀꢄ ꢁꢂꢃ) by the Hypo 4 model for test
set-1 and test set-2 were calculated and found to be 0.93 and 0.61, respectively. The rꢄ
ꢀ
ꢁꢂꢃ
value (0.93) of test set-1 is higher than the correlation value (0.88) of the training set for
hypothesis 4. This was due to the presence of an outlier compound 37 in the test set-1, which
had an IC50 of 10,478 nM. This value is over the maximum predictive ability of the training
set; the maximum IC50 of the training set compound was 7060 nM (for compound 20). When
11

ACCEPTED MANUSCRIPT
the outlier compound 37 was removed, the r ꢀꢄ ꢁꢂꢃꢅvalue of test set-1 by Hypo 4 decreased to
0.86. Also the predictive value of Hypo 1, 2, and 3 for test set-1 was calculated as 0.61, 0.46
and 0.37, respectively (Table S6, supplemental information), which was much lower than that
of Hypo 4. These results suggest that the model Hypo 4 is able to predict the aurora kinase A
inhibitory potential of both the training set related as well as unrelated compounds, with
reasonable accuracy.
3
.2. Development & validation of ligand efficiency based virtual screening strategy
In the second step, the pharmacophore model was used to screen in-house HTS library of
25,000 compounds to identify novel aurora kinase inhibitors. This could be accomplished by
1
predicting the aurora kinase A activity (IC50) and then ranking (PH rank) the database
compounds using the model Hypo 4 and then selecting the top ranked compounds for
biochemical testing. In spite of several successful examples of hits identified from VS
methods, people are still working to improve the hit rate of VS in order to identify hits cost
and resource efficiently. Hit rates refer to the percent of positive hits identified during a
database screening campaign; this varies widely, depending upon several factors, including
the screening method (SBVS or LBVS) and database used (general library or focused library).
Our aim is to efficiently identify aurora kinase lead compounds from a database by utilizing a
LE based VS platform.
For this purpose we planned to use ligand efficiency (LE) along with the
12

ACCEPTED MANUSCRIPT
pharmacophore based scoring, so as to identify aurora kinase inhibitors more efficiently.
Many groups have now used LE metrics to assess hits from VS as well as use the metrics for
optimizing hits in drug discovery programs [3, 16]. The LE metric was first defined by
∆
ꢆ
ꢀꢉꢊꢋꢌ
Andrews et al. [16], which defines it as biological activity per molecular size ( ꢅorꢅ ꢇꢈ
;
ꢇꢈ
where ∆G = free energy of binding, HA = number of non hydrogen heavy atoms). Although
LE is currently an useful tool in fragment-based drug design (FBDD), it cannot be solely used
to judge the better of two fragments with disparate size, since the LE range varies depending
on the molecular size of the fragments under consideration [38]. In order to overcome the
drawbacks of the LE function and also as an improvement to LE, Reynolds et al., compared
more than 8000 ligands and their 28 protein targets and defined a new function, the “fit
quality” (FQ) score, to remove the weight dependent problem of ligand efficiency [39]. The
fit quality (FQ) score can be defined as size normalized LE metric and can be calculated from
the following equation (1):
ꢎꢏ
FQ ꢍ
(1)
ꢎꢏ_ꢐꢑꢒꢓꢂ
where LE is the ligand efficiency and LE_Scale is derived from the top ligand efficiency
versus heavy atom count curve. Moreover, different types of targets will have different
LE_Scale distributions. Hence, to use FQ as a part of the VS scoring function, the LE_Scale
for aurora kinase A target needs to be calculated over a range of heavy atom counts. So a
diverse set of aurora kinase inhibitors were collected and used to generate the LE_Scale
13

ACCEPTED MANUSCRIPT
function based on these compounds.
This diverse set contains 234 known aurora kinase inhibitors with seven different
cores (Group 1-7) [40] and 9 aurora kinase inhibitors in clinical trials [29-37] (total 243
control compounds with aurora kinase A IC50 ranging from 1 to 17,219 nM and HA count
i
ranging from 16 to 55). The data (IC50 or K values) were collected from literature and their
i
LE (pIC50/HA or pK /HA) were calculated (Table S7, supplemental information). In order to
derive the LE_Scale, the plot of LE against HA count was generated for all the 243 aurora
kinase inhibitors (Figure 2). The LE_scale function was derived by fitting a curve to the
highest LE value over the whole range of HA count using OriginPro 7.0 software (OriginLab,
Northampton, MA). The equation (2) describes the fitted curve representing the LE_Scale for
aurora kinase A target.
2
3
4
LE_Scale = 0.86824-(0.0563)*HA+0.00237*HA -(498971E-5)*HA +(38647E-7)*HA (2)
Once LE_Scale function was derived over a range of HA count for the aurora kinase
A target, database compounds identified by the pharmacophore model (PH rank) could be FQ
scored using equation -1 and reranked (PH_FQ rank). Based on the PH_FQ rank order,
compounds could be biochemically tested for their aurora kinase A inhibition. However,
randomly choosing the top 100 or 200 compounds from PH_FQ screening may not give
optimal results. Hence, to identify the optimal screening threshold point, we planned to
calculate the enrichment factor (EF) at different screening threshold points based on the
14

ACCEPTED MANUSCRIPT
recovery of 243 known aurora kinase inhibitors from the database by both PH and PH_FQ
ranking.
Applying the above strategy, 125,243 compounds (125,000 database + 243 control
compounds), were subjected to pharmacophore screening. Out of those, only 894 compounds
could be matched to the pharmacophore feature of Hypo 4, with a maximum of 2 features
omitted. Based on the predicted IC50 values, these compounds were ranked (PH rank). Then,
LE values were calculated using the predicted IC50 for the 894 compounds and the LE_Scale
value for their HA count derived from equation (2). FQ values were computed for these
compounds from the LE and LE_Scale values using equation (1). Based on the FQ values,
these 894 compounds were reranked (PH_FQ rank). All 894 compound codes, along with
their predicted IC50, LE, LE_Scale, PH rank, and PH_FQ rank, are shown in Table S8,
supplemental information.
Next, the enrichment factor (EF) for the screening was calculated based on the
retrieval of control compounds from the database using the equation (3) [41], at different
screening threshold levels of 1 to 30%:
ꢞ%
ꢞ%
ꢇꢕꢖꢗ
/ꢟ
ꢘꢙꢚꢙꢛꢜꢙꢝ
EFꢔ%
ꢍ
ꢘꢙꢚꢙꢛꢜꢙꢝ
(3)
ꢇꢕꢖꢗꢜꢠꢜꢡꢚ/ꢟꢜꢠꢜꢡꢚ
where Hitsꢔ%
is the number of active ligands present in the top x% of the
ꢗꢂꢓꢂꢑꢖꢂꢃ
database, Hitsꢖꢢꢖꢒꢓ is the total active ligands in the database, Nꢔ%
is number of the
ꢗꢂꢓꢂꢑꢖꢂꢃ
compounds in the top x% of the database, Nꢖꢢꢖꢒꢓ is number of compounds in the entire
15

ACCEPTED MANUSCRIPT
1
%
database. EF means enrichment of hits obtained if the top 1% of the database is screened. A
1
%
random screening of the database would given an EF = 1 at the 1% threshold level. Even if
the VS protocol can improve EF by a few fold, the method could be considered to be good
since the chance of identifying a hit is improved, compared to random screening.
EF for both PH and PH_FQ screening methods at different threshold levels of
screening are shown for control compounds in Figure 3A and in Table S9, supplemental
information. It shows that incorporation of LE based scoring into the screening protocol
improves the hit identification rate (higher EF) at all levels of screening threshold, compared
to PH based screening alone. For example, at a screening threshold of 25%, EF for PH based
screening is only 2.61; while for PH_FQ based screening EF is 3.48 (Table S9, supplemental
information). As maximum EF for PH_FQ based screening was observed at the 25%
screening threshold, it was chosen as the threshold point for selecting database compounds
for biochemical testing.
To further confirm that PH_FQ based ranking would improve the chance of
identifying the hits better than just using PH based ranking, we analyzed the retrieval of
control compounds by both methods using receiver operating characteristic (ROC) curve [41,
42]. ROC curve helps to judge the performance of a binary classifier system and provides
tools to select optimal models and to discard suboptimal ones. ROC curve was constructed by
plotting a graph of sensitivity (Se, true positive rate) vs. 1-specificity (1-Sp, false positive rate)
16

ACCEPTED MANUSCRIPT
for both PH and PH_FQ based screening methods. The measure of Se and Sp are defined as
ꢣꢤ
Se ꢍ
Sp ꢍ
(4)
(5)
ꢣꢤꢥꢦꢟ
ꢣꢟ
ꢣꢟꢥꢦꢤ
where TP is the number of correctly identified active ligands (true positives), TN the number
of correctly identified inactive ligands (true negatives), FP the number of incorrectly
identified active ligands (false positives), and FN the number of incorrectly identified
inactive ligands (false negatives). The area under the ROC curve (AUC) is a useful measure
of the performance of two different virtual screening approaches. The ideal screening method
will have an AUC value of 1, while a random screening method would lead to an AUC value
of 0.5. The AUC values of the PH and PH_FQ screening methods are 0.65 and 0.84,
respectively (Figure 3B). ROC curve analysis further reiterates that the PH_FQ based
screening method is better than PH based screening alone and can improve the chance of
identifying aurora kinase inhibitors from the database.
3.3. Virtual screening and hit identification
As a third step, based on the PH_FQ reranking, top 223 compounds (25% screening threshold)
out of 894 compounds were selected, out of which 72 were control compounds and remaining
151 were database compounds. The top ranked 151 database compounds were tested at 10
µM concentration for aurora kinase A inhibition using kinase glow assay (Table S8,
supplemental information). Out of the 151 compounds tested, 7 compounds, 1-7 showed
17

ACCEPTED MANUSCRIPT
more than 50% inhibition at 10 ꢀM (Table 1). All seven compounds were subjected to IC50
determination, which showed that 5 has an IC50 of 1.29 µM. In order to confirm that 5 is a
potent aurora kinase A inhibitor and also to assess its overall kinase profile we synthesized
the compound as shown in scheme 2 and tested the compound in a panel of 31 kinases at 10
µM concentration using Life Technologies Corporation, Z-LYTE® screening assay (Table
S10, supporting information). The kinase profiling data demonstrate that compound 5 is a
potent and selective aurora kinase inhibitor.
Further, the lead 5 identified was docked to aurora kinase A crystal structure (PDB ID:
2W1C) and mapped to pharmacophore Hypo 4, with an aim to understand the protein-ligand
interactions (Figure 4). Docking showed that the phenolic OH group acts as a hydrogen bond
acceptor (HBA) to interact with the Glu211 residue in the hinge region and the NH group of
amide functionality acts as a hydrogen bond donor (HBD) to interact with Glu181 residue in
the αC helix. An additional H-bond was observed between the Asp274 of the DFG loop and
the amide carbonyl group next to the furan ring. The furan ring maps onto the HY3
hydrophobic region and makes hydrophobic contact with Glu181, Asp274, and Lys162. The
two phenyl rings (Cl-Ph- and =N-Ph-NH-) map to the HY1 and HY2 regions of the
pharmacophore model and makes extensive contact with the Leu 139, Gly 140, Lys141, Val
147, Lys162, Glu181, Leu210, and Asp274 residues in the ATP binding site.
To confirm that the PH_FQ screening had actually improved the hit rates, the rank
18

ACCEPTED MANUSCRIPT
order of active hits in each screening methods can be compared (Table 1). When the
screening threshold is set to top 100 compounds, there is 0 hits from PH screening method
and 3 hits from PH_FQ screening method. If the screening threshold is set as 214 compounds,
PH_FQ screening method would have identified all the 7 hits, but PH screening would have
only identified 2 hits. Moreover, comparison of EF for VS method with and without FQ
ranking for the identification of control as well as new hits (from database) shows that the VS
incorporating FQ based scoring is better in both the cases (Table 2). This further emphasize
that incorporation of the LE based metric, FQ, to the pharmacophore based screening had
improved the hit rate of VS and helped in identifying a selective aurora kinase inhibitor 5.
4. Conclusion
Identification of a suitable hit molecule for further optimization to a lead is an important step
in the early stages of drug discovery. In this regard, nowadays virtual screening (VS) is
extensively used to identify suitable hits and methods to improve the VS hit rate (% of hits
identified to compounds tested biochemically) is much sought after. In this study, we report
for the first time the use of a ligand efficiency (LE) based metric, ‒ Fit quality (FQ), to rank
database compounds and prioritize them for biochemical testing in aurora kinase A inhibition
assay.
The in-house database compounds (125,000) were first screened using a
pharmacophore model (PH screening) developed using 20 known aurora kinase inhibitors.
19

ACCEPTED MANUSCRIPT
Next, LE_Scale function was derived from 294 known aurora kinase inhibitors, which
defines the maximum LE achievable for a particular heavy atom count for the aurora kinase
target. The LE_scale function was used to calculate the fit quality (FQ) score for PH screened
database compounds. Based on the FQ score (PH_FQ screening), the top 151 compounds
(0.12% of database) were tested biochemically to identify 5 as a novel aurora kinase selective
inhibitor. Incorporating a LE based metric, FQ, into VS clearly improved the enrichment
factor (EF) — from 237 (PH screening) to 828 (PH_FQ screening) — indicating that FQ
based VS is more efficient method for hit identification. The LE based VS protocol described
here for aurora kinase hit identification could be utilized for other drug targets to identify hits
more efficiently.
5. Experimental
5.1. Pharmacophore generation
The pyrazole and furanopyrimidine aurora kinase inhibitors reported earlier were assigned to
training and test set compounds as described before (Table S1 and S2, supplemental
information) [27]. The pharmacophore generation was performed using Discovery
Studio/3D-QSAR pharmacophore generation software (Discovery Studio 2.1, Accelrys Inc.,
San Diego, CA, USA). In the parameter setting, the Uncert value was set to 1.4, due to the
narrow range of activities (IC50 values of 33–7060 nM) of the compounds. Minimum
interfeature distance was set as 2.5 Å, and the maximum excluded volumes were set to 100.
20

ACCEPTED MANUSCRIPT
Minimum and maximum pharmacophore features were set as 4 and 5, respectively. Three
pharmacophore features were derived by the HypoGen algorithm: hydrogen bond acceptor
(HBA), hydrogen bond donor (HBD), and hydrophobic (HY). The quality of the
pharmacophore hypothesis constructed was assessed by the Fischer validation, with
confidence level of 95%, and was based on Fischer’s randomization test by randomly
assigning the activities of the training set molecules as new spreadsheets. The prediction
ability of the pharmacophore was evaluated by the goodness of prediction of activity for 55
test set aurora kinase inhibitors (Test set - 1: 46 pyrazoles and furanopyrimidines, Table S2,
supplemental information + Test set - 2: 9 aurora kinase inhibitors in clinical trials, Table S5,
supplemental information). The linear regression coefficient for the activity prediction (r ꢀꢄ ꢁꢂꢃ)
was used as a parameter to ascertain the prediction ability of the model.
5.2. Virtual screening
The first round of VS was done by using Discovery Studio/ Ligand Pharmacophore Mapping
program. The Hypo 4 was used for screening the in-house database (125,243 = 125,000
database + 243 control compounds). The database compound conformations were generated
by Discovery Studio/ Conformation program; the “best” method and the conformers were
created within the relative energy threshold of 20 kcal/mol. The maximum number of
conformations allowed for each compound was set to 300. The conformations generated were
used for performing the virtual screening. The screening method used the “best mapping”
21

ACCEPTED MANUSCRIPT
with flexible fit between ligand and the pharmacophore. The maximum number of features
that are allowed to miss when mapping the ligand to the pharmacophore was set to 2. The
minimum distance between ligand and features was set as 2 Å. Pharmacophore based ranking
(PH rank) of the database compounds using Hypo 4 resulted in the identification of 894
compounds (811 database + 83 control compounds) from the database. The second run of VS
was done by LE based screening. The heavy atoms of the 894 compounds were calculated
and used for determining the FQ value of the compounds (FQ = LE/LE_scale). Finally, the
894 compounds were ranked (PH_FQ rank) by their FQ value and the top 25% (223 = 151
database + 72 control compounds) database compounds sent for biochemical testing.
5.3. Synthesis of hit 5
5.3.1. 3-Bromo-5-[(2-chloro-phenyl)-hydroxy-methyl]-2-hydroxy-benzoic acid methyl ester
(10).
A solution of compound 9 (800 mg, 3.09 mmol) in THF (15 mL) was added to 2-
chlorophenylmagnesium bromide (9.27 mmol) in THF (9 mL) and stirred for 2 h under ice
4
bath condition. The reaction was quenched by the addition of saturated NH Cl (aq.) (20 mL),
followed by extraction with EtOAc (2 × 20 mL). The combined organics were dried over
4
MgSO and concentrated under reduced pressure. The residue obtained was purified by flash
column chromatography on silica gel (EtOAc/Hexane = 1:4) to furnish compound 10 (741.4
1
mg, 65%). H NMR (400 MHz, CDCl
3
): δ 11.45 (s, 1H), 7.88 (d, J = 2.0 Hz, 1H), 7.74 (d, J =
22

ACCEPTED MANUSCRIPT
2
1
1
.0 Hz, 1H), 7.58 (d, J = 7.6 Hz, 1H), 7.38-7.26 (m, 3H), 6.16 (s, 1H), 3.96 (s, 3H), 2.41 (s,
13
3
H). C NMR (100 MHz, CDCl ): δ 169.7 (C), 157.0 (C), 139.9 (C), 137.1 (CH), 134.1 (C),
31.8 (C), 129.3 (CH), 128.7 (CH), 127.4 (CH), 127.2 (CH), 127.0 (CH), 112.7 (C), 110.9
+
(
C), 70.8 (CH), 52.6 (CH
3
). LCMS (ESI) m/z: 394.9 (M+Na) .
5.3.2. 3-Bromo-5-(2-chloro-benzyl)-2-hydroxy-benzoic acid methyl ester (11).
2 2
To a solution of compound 10 (741.4 mg, 2.08 mmol) in CH Cl (10 mL), trifluoroacetic acid
(1.14 g, 10.00 mmol) and triethylsilane (697.7 mg, 6.00 mmol) were added, and the mixture
was stirred for 1 h at room temperature. Then the reaction mixture was concentrated under
reduced pressure to furnish compound 11 (615.3 mg, 83%), which was used for the next step
1
without further purification. H NMR (400 MHz, CDCl
3
): δ 11.33 (s, 1H), 7.66 (d, J = 2.4 Hz,
1H), 7.55 (d, J = 2.4 Hz, 1H), 7.39 (d, J = 6.8 Hz, 1H), 7.22-7.13 (m, 3H), 4.01 (s, 2H), 3.95
13
(
3
s, 3H). C NMR (100 MHz, CDCl ): δ 170.1 (C), 156.6 (C), 139.4 (CH), 137.7 (C), 134.1
(C), 131.3 (C), 130.8 (CH), 129.5 (CH), 129.4 (CH), 128.1 (CH), 127.0 (CH), 113.2 (C),
+
1
11.2 (C), 52.7 (CH
3
), 38.0 (CH
2
). LCMS (ESI) m/z: 378.9 (M+Na) .
5.3.3. 2-Bromo-4-(2-chloro-benzyl)-6-hydroxymethyl-phenol (12).
To a solution of lithium aluminum hydride (124.5 mg, 3.28 mmol) in THF (2 mL) maintained
under ice bath condition, a solution of compound 11 (584.5 mg, 1.64 mmol) in THF (2 mL)
2 4
was added and stirred for 2 h. The reaction was quenched by the addition of saturated Na SO
(
2
aq.) (0.3 mL) and stirring the mixture for 10 min. Then the mixture was extracted with Et O (2
23

ACCEPTED MANUSCRIPT
×
2
20 mL), and the combined organics were washed with H O and brine. Then the organics
were dried over MgSO
4
and concentrated under reduced pressure to furnish compound 12
1
(
532.5 mg, quant.), which was used for next step without further purification. H NMR (300
3
MHz, CDCl ): δ 7.38 (d, J = 5.1 Hz, 1H), 7.36 -7.15 (m, 4H), 6.97 (s, 1H), 4.75 (s, 2H), 4.00
13
(
3
s, 2H). C NMR (100 MHz, CDCl ): δ 149.6 (C), 138.1 (C), 134.1 (C), 132.8 (C), 131.9
(CH), 130.9 (CH), 129.7 (CH), 128.2 (CH), 127.9 (CH), 127.1 (C), 127.0 (CH), 110.4 (C),
+
63.0 (CH
2
), 38.0 (CH
2
). LCMS (ESI) m/z: 350.9 (M+Na) .
5.3.4. 3-Bromo-5-(2-chloro-benzyl)-2-hydroxy-benzaldehyde (13).
2 2
To a solution of compound 12 (532.5 mg, 1.63 mmol) in CH Cl (8 mL) and MeOH (0.8 mL),
2
MnO (1.42 g, 16.3 mmol) was added and stirred for 8 h under room temperature. The
mixture was concentrated under reduced pressure and purified by flash column
chromatography on silica gel (EtOAc/Hexane = 1:2) to furnish compound 13 (408.1 mg,
1
7
7%). H NMR (400 MHz, CDCl
3
): δ 11.47 (s, 1H), 9.79 (s, 1H), 7.64 (d, J = 2.0 Hz, 1H),
13
7
.41 (d, J = 6.4 Hz, 1H), 7.32 (d, J = 2.0 Hz, 1H), 7.25-7.19 (m, 3H), 4.08 (s, 2H). C NMR
3
(100 MHz, CDCl ): δ 195.8 (CH), 156.3 (C), 140.2 (CH), 140.1 (C), 137.1 (C), 134.0 (CH),
2
132.6 (C), 130.8 (CH), 129.7 (CH), 128.2 (CH), 127.1 (CH), 120.8 (C), 111.0 (C), 37.7 (CH ).
+
LCMS (ESI) m/z: 348.8 (M+Na) .
5.3.5. Furan-2-carboxylic acid (4-nitro-phenyl)-amide (15).
To a solution of 4- nitroaniline (500 mg, 3.62 mmol) and triethylamine (54.9 mg, 0.54 mmol)
24

ACCEPTED MANUSCRIPT
in CH
2
Cl
2
(18 mL) was added 2-furoyl chloride (472.5 mg, 3.62 mmol), and the mixture was
stirred for 8 h at room temperature. After completion of the reaction, the organics were
2 4
washed with H O (2 × 10 mL). The organic layer was dried over MgSO and concentrated
under reduced pressure. The residue was purified by flash column chromatography on silica
1
gel (EtOAc/Hexane = 1:3) to furnish compound 15 (622.4 mg, 74%). H NMR (400 MHz,
3
CDCl ) δ 8.35 (s,1H), 8.27 (d, J = 7.2 Hz, 2H), 7.85 (d, J = 7.2 Hz, 2H), 7.57 (d, J = 1.6 Hz,
1
3
1
1
3
H), 7.33 (dd, J = 3.6, 1H), 6.62 (dd, J = 3.6, 1.6 Hz, 1H). C NMR (75 MHz, CD OD): δ
48.3 (C), 147.5 (C), 143.4 (C), 127.1 (CH ×2), 125.8 (CH), 121.4 (C), 119.4 (CH ×2), 117.4
+
(
CH), 113.6 (CH). LCMS (ESI) m/z: 255.0 (M+Na) .
5.3.6. Furan-2-carboxylic acid (4-amino-phenyl)-amide (16).
A solution of compound 15 (498.9 mg, 2.15 mmol) and 10% Pd/C (50 mg) in EtOH (10 mL)
was stirred under hydrogen atmosphere for 8 h. The reaction mixture was filtered over Celite,
and the solvents removed by reduced pressure. The residue obtained was purified by flash
2 2
column chromatography on silica gel (MeOH/CH Cl = 1:40) to furnish compound 16 (413.8
1
mg, 95%). H NMR (300 MHz, CDCl
3
) δ 7.92 (s, 1H), 7.49 (dd, J = 1.8, 0.9 Hz, 1H), 7.42 (d,
J = 8.7 Hz, 2H), 7.20 (d, J = 3.3 Hz, 1H), 6.69 (d, J = 8.7 Hz, 2H), 6.54 (dd, J = 3.3, 1.8 Hz,
13
1
3
H), 3.64 (s, 2H). C NMR (100 MHz, CDCl ): δ 155.9 (C), 148.0 (C), 143.9 (CH), 143.5
(C), 128.6 (C), 121.9 (CH ×2), 115.4 (CH ×2), 114.7 (CH). LCMS (APCI) m/z: 203.1
+
(
M+H) .
25

ACCEPTED MANUSCRIPT
5.3.7. Furan-2-carboxylic acid (4-{[3-bromo-5-(2-chloro-benzyl)-2-hydroxy-benzylidene] -
amino}- phenyl)-amide (5).
A solution of compound 13 (179.1 mg, 0.55 mmol) and compound 16 (110.7 mg, 0.55 mmol)
in EtOH (3 mL) was stirred under reflux for 2 h. The precipitate formed was filtered to
1
furnish compound 5 (145.4 mg, 52%). H NMR (400 MHz, d
6
-DMSO): δ 10.33 (s, 1H), 8.97
(s, 1H), 7.95 (d, J = 1.6 Hz, 1H), 7.85 (d, J = 8.8 Hz, 2H), 7.56 (d, J = 2.0 Hz, 1H), 7.50 (d, J
13
=
8.8 Hz, 2H), 7.46-7.26 (m, 6H), 6.71 (dd, J = 3.6, 1.6 Hz, 1H), 4.05 (s, 2H). C NMR (100
6
MHz, d -DMSO): δ 161.4 (CH), 156.2 (C), 156.0 (C), 147.4 (C), 145.8 (CH), 141.7 (C),
1
1
1
38.2 (C), 138.0 (C), 135.9 (CH), 133.1 (C), 131.9 (CH), 131.4 (CH), 131.1 (C), 129.5 (CH),
28.4 (CH), 127.5 (CH), 122.0 (CH × 2), 121.0 (CH × 2), 119.7 (C), 115.0 (CH), 112.2 (CH),
+
09.9 (C), 37.0 (CH
2
). LCMS (APCI) m/z: 511.0 (M+H) .
5.4 Aurora kinase A inhibition assay
Aurora kinase A inhibition assays for the compounds were carried out as described by us
previously [43].
Supplementary data
Supplementary data associated with this article can be found in the online version. These data
include the training set and test set of the pharmacophoe model, statistical parameters and
features of the top ten hypotheses, known aurora kinase inhibitors used for deriving FQ, FQ
rank calculation, and enrichment factor (EF) calculation, as well as the kinase selectivity
26

ACCEPTED MANUSCRIPT
1
13
profile of 5 and H NMR and C NMR scan of synthesized compounds.
Abbreviations
3D-QSAR: three dimensional quantitative structure-activity relationship, EF: enrichment
factor, FBDD: fragment-based drug discovery, FQ: fit quality, HA: number of heavy atoms,
HBA: hydrogen bond acceptor, HBD: hydrogen bond donor, HTS: high through-put
screening, HY: hydrophobic, LBVS: ligand-based virtual screening, LE: ligand efficiency, PH:
pharmacophore model, ROC: receiver operating characteristic, SBVS: structure-based virtual
screening, VS: virtual screening.
Acknowledgements
Financial support from the National Health Research Institutes and the National Science
Council, Taiwan (NSC-102-2325-B-400-003 and NSC-101-2113-M-400-002-MY4 to HPH;
NSC-102-2113-M-400-003-MY3, to HYS), and from the Science & Engineering Research
Board, India (SR/FT/LS-64/2011 to MSC), are gratefully acknowledged.
27