Communication www.molinf.
com
DOI: 10.1002/minf.201300072
Identification of Novel Histamine H4 Ligands by Virtual
Screening on Molecular Dynamics Ensembles
Rbert Kiss,[a] Balzs Jjrt,[b] va Schmidt,[c] Bla Kiss,[c] and Gyçrgy M. Keserű*[c, d, e]
Abstract: We report the identification of novel histamine
H4 receptor ligands by ensemble docking on homology
model conformers derived from molecular dynamics simu-
lations. Selected receptor models from the trajectories
demonstrated superior virtual screening performance com-
pared to the initial models. The ensemble of the best
Keywords: Histamine H4 · Structure-based virtual screening · Homology model · Molecular dynamics · Ensemble docking
Atomic resolution structure of drug targets is usually a pre-
requisite for structure-based drug design. In the absence of
experimental structures discovery projects can utilize ho-
mology models. Although comparative modeling has cer-
tain limitations, homology models have been proven to be
useful in early phase drug discovery.[1] In particular, carefully
designed virtual screens with optimized docking protocols
have been successfully conducted on homology models
and are generally used for the identification of chemical
starting points.[2–5] On the other hand, however, high
throughput docking algorithms typically do not take into
account protein flexibility. The application of a single con-
formation for structure-based virtual screening is therefore
suboptimal[6–8] especially in the case of homology models
where the uncertainty of the side chain orientations is con-
siderably high. It is therefore advisable to run the virtual
screen on an ensemble of protein conformations with low
abundance of high energy protein conformations to reduce
noise.[8] One viable strategy might involve the analysis of
numerous individual conformations and the selection of
those conformations showing the highest efficiency and di-
versity of retrieved actives. Conformational ensembles of
proteins can be generated in various ways, such as by ap-
plying Monte Carlo or molecular dynamics (MD) simula-
tions[9] or normal-mode analysis.[10]
Figure 1. Structure of histamine (1) and JNJ-7777120 (2).
 2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
models was able to retrieve a diverse set of known H4 li-
gands. Prospective virtual screening against these models
and subsequent in vitro experimental validation identified
novel H4 ligands. Compound 3 showing highest affinity
and ligand efficiency represents an interesting scaffold for
further medicinal chemistry exploration.
The trajectories from the previously published MD simu-
lations of human histamine H4 receptor (hH4R) models
generated in the presence of the endogenous activator his-
tamine (1) and the selective H4 antagonist JNJ-7777120 (2)
(see compounds 1 and 2 in Figure 1, respectively)[11] provid-
ed us a unique source of conformer ensembles that were
evaluated in prospective virtual screens.
In the referred MD study we found characteristics of
GPCR activation in the histamine-hH4R simulation, while
the simulation in the presence of JNJ-7777120 seemed to
be tailored towards the inactive state of hH4R.[11]
In this study we report molecular dynamics (MD) based
ensemble docking used for the identification of novel H4 li-
gands by structure-based virtual screening. Our retrospec-
[a] R. Kiss
mcule.com Ltd.
Vendel utca 15–17, B/2/6, H-1096 Budapest, Hungary
[b] B. Jjrt
Department of Chemical Informatics, Faculty of Education,
University of Szeged
Boldogasszony sgt. 6., H-6725, Szeged, Hungary
[c] . Schmidt, B. Kiss, G. M. Keserű
Gedeon Richter Plc
Gyçmrői fflt 19–21., H-1103, Budapest, Hungary
*e-mail: gy.keseru@ttk.mta.hu
[d] G. M. Keserű
Department of General and Analytical Chemistry, Budapest
University of Technology and Economics
Szt. Gellrt tr 4., H-1111, Budapest, Hungary
[e] G. M. Keserű
Present address: Research Centre for Natural Sciences, Hungarian
Academy of Sciences
Pusztaszeri fflt 59–67., H-1025 Budapest, Hungary
Supporting Information for this article is available on the WWW
under http://dx.doi.org/10.1002/minf.201300072.
Mol. Inf. 2014, 33, 264 – 268 264
Communication www.molinf.com

tive screens revealed that certain receptor conformers were
more efficient in retrieving H4 antagonists than the initial
models. Based on these results, we limited our prospective
screen to an ensemble of the most efficient models. Pro-
spective screening on the selected models identified novel
H4 scaffolds suitable for further exploration.
First we tested the performance of the individual snap-
shots obtained from our MD simulations performed on the
histamine-hH4R and the JNJ-7777120-hH4R complexes. The
results showed that the histamine-hH4R models were
better suited for retrieving agonists than the JNJ-7777120-
hH4R models. On the other hand, snapshots taken from
the JNJ-7777120-hH4R simulation could identify antago-
nists with higher efficiency than those from the histamine-
hH4R one. This is in line with several observations de-
scribed in Jojrt et al.[11] showing GPCR activation character-
istics of the histamine-hH4R system, while no such observa-
tions were obtained for the JNJ-7777120-hH4R complex
which showed characteristics of inactive state of GPCRs. We
analyzed the reasons behind this major difference between
the two complexes in more detail.
H4 agonists are typically less complex compounds with
lower molecular weight than H4 antagonists. It was expect-
ed that this difference in size plays a role in the recognition
of agonists vs. antagonists by the hH4R models. In particu-
lar, receptor models optimized in the presence of the small
molecule agonist histamine might not be able to accom-
modate larger ligands. This hypothesis is supported by the
fact that the number of successfully docked H4 antagonists
and corresponding decoys (larger ligands) was 13 % lower
for the histamine-hH4R models than for the JNJ-7777120-
hH4R ones, while the difference was only 4 % for the H4 ag-
onists and their corresponding decoys (smaller ligands).
The high number of unsuccessful dockings is typically
a consequence of steric clashes between the ligands and
the receptor. According to these findings it seems that JNJ-
7777120-hH4R models are better suited to identify novel
H4 antagonists.
We have investigated whether MD simulation of the his-
tamine-hH4R complex yielded more suitable models for ag-
onist identification than the initial complex. We found,
however, that the initial structure was the most efficient in
H4 agonist identification. This is in line with our previous
findings showing that the initial histamine-hH4R model al-
ready shows some characteristics of the active state (e.g.
missing ionic lock). This might be reasoned by the fact that
hH4R has a significant constitutive activity.
We compared the enrichment factors achieved by the
JNJ-7777120-hH4R snapshots to analyze the screening per-
formance as a function of simulation time. We have ana-
lyzed the enrichment factors at 2–5 % of the retrospective
databases ranked by the docking scores (Figure 2), which

Figure 2. Enrichment factors obtained with the rhodopsin-based (A) and H1R-based (B) JNJ-7777120-hH4R models as a function of MD
simulation time. Top 2 % (black, solid line), 3 % (black, dashed line), 4 % (grey, solid line) and 5 % (grey, dashed line) of the ranked databases
are shown.

 2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim Mol. Inf. 2014, 33, 264 – 268 265
Communication
seemed to be a reasonable range for hit selection in the
planned prospective screens (see below). Besides of the
rhodopsin-based H4R models[11] we conducted enrichment
tests on the snapshots of another MD simulation conduct-
ed on the recently published histamine H1 receptor (H1R)
based homology model.[12] Enrichment factors retrieved in
different ranges and also for the published GLL and GDD
ligand and decoy sets[13] can be found in the Supporting In-
formation.
Interestingly we found that several snapshots showed su-
perior performance compared to the initial structure. First,
we analyzed the performance of the rhodopsin-based H4R
models. Considering that a crucial salt bridge between JNJ-
7777120 and Asp94(3.32) is stabilized after 9.2 ns of simula-
tion,[11] we further focused our selection on models after
this interaction was established. Snapshots taken at 14 ns
(Model A), 16 ns (Model B) and 17 ns (Model C) yielded en-
richment factors 9.5, 7.1 and 7.1, respectively at the top 2 %
of the ranked database. We compared the structures of the
actives identified by these models and found that all ac-
tives identified by Model B could be also retrieved by the
combination of Models A and C (Figure 3).
On the other hand, Models A and C identified different
actives. This suggested that a combined approach by
screening on an ensemble of the best performed models
could identify diverse H4 scaffolds. We therefore selected
Models A and C for prospective screening for identifying
novel H4 antagonists.
We also analyzed the performance of the H1R based
models and selected snapshots taken at 16 ns and 20 ns
for further inspection. The actives retrieved at the top 2 %
of the ranked database by either of these models were
Figure 3. H4 reference ligands retrieved by Models A (left), B (middle) and C (right) in the retrospective tests.
 2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
www.molinf.com
identical. We therefore selected the snapshot taken at
16 ns for prospective screening (Model D).
A prefiltered subset of our in-house compound database
has been screened against rhodopsin-based Models A and
C. Altogether 2193 compounds were docked into the bind-
ing site of both models and the top 2 % of each ranking
lists were selected for in vitro testing. Due to the overlap
between the two hit lists, 75 compounds have been sub-
jected to experimental testing. Seven compounds exhibited
an inhibition above 30 % at 5 mM. Four of these samples
had purity above 90 % (Figure 4). H4 Ki values of these com-
pounds have been determined and ligand efficiency (LE)[14]
and ligand efficiency dependent lipophilicity (LELP)[15]
values have been calculated.
Four actives out of the 75 tested compounds correspond
to a hit rate of 5.3 %. Thirteen out of the 75 tested com-
pounds were top ranked by both Models A and C. Two of
these 13 compounds were found to be among the actives.
This consensus scoring approach yielded a higher hit rate
(15.4 %). These hit rates are similar to that achieved in our
previously published large-scale virtual screen on hH4R
(6.3 %). It has to be mentioned that these new actives have
been found by screening only a small portion of our in-
house database, while our previously published large-scale
virtual screen was conducted on more than 8 million mole-
cules. Accordingly, the performance of the models derived
from the MD simulation is remarkable. Tanrikulu et al. previ-
ously published a virtual screen on pseudoreceptor models
derived from MD simulations.[16] In their study, two out of
the nine tested compounds were found to possess moder-
ate H4 affinity (pKi = 4.5) (hit rate: 22.2 %).
Considering both LE and LELP values the most valuable
hit is compound 3 with a Ki of 4.4  0.4 mM. It is also worth
Mol. Inf. 2014, 33, 264 – 268 266
Communication www.molinf.com

Figure 4. H4 hits identified by structure-based virtual screening on Models A (4) or C (6) or both (3 and 5) and Model D (7). Ki values are
calculated as the results of at least three experiments.

mentioning that compound 3 has been top ranked by
both models A and C. While the piperazine group of com-
pound 3 is commonly found in previously published H4 li-
gands, the thiocarbamide core makes it an interesting,
novel starting point for further exploration. The predicted
binding conformation of compound 3 strongly overlaps
with that of JNJ-7777120 and it forms critical interactions
with Asp94(3.32) and Glu182(5.46) (see Figure S2 in Sup-
porting Information). Compounds 4 and 5 share the carba-
zole core. These compounds have suboptimal LE and LELP
values, but they might be improved by modifying their
sidechains.
Additionally, we performed a prospective screen on the
H1R-based Model D. The same dataset was screened
against this model and the top 2 % (44 compounds) was se-
lected for further investigation. Twenty-three compounds
had purity above 90 % that were subjected to experimental
testing. We identified 1 hit (compound 7) that showed an
inhibition above 30 % at 5 mM. This corresponds to a hit
rate of 4.3 %, which is similar to that found with the rho-
dopsin-based models.
All hits were subjected to a novelty check at the ChEMBL
database.[17] Similarity searches were performed by using
the hits as queries at the lowest available cutoff (70 %). Bio-
activity data for all retrieved analogs were analyzed but no
data on their histamine H4 (or other subtype) affinity could
be found.

 2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim Mol. Inf. 2014, 33, 264 – 268 267
Communication www.molinf.com
In summary, here we analyzed the impact of MD simula-
tion and ensemble docking on homology model based vir-
tual screening. The enrichment tests showed that some
snapshots derived from MD simulation of the hH4R-
JNJ7777120 complex were more suitable for the identifica-
tion of H4 antagonists than the initial models. Additionally,
we found that MD simulation of a homology model based
on the rhodopsin template resulted in models with similar
docking predictive power as the ones based on the H1R
template. This suggests that MD simulations can sample
the conformation space of H4R homology models based
on either of these templates. Ensemble docking on the se-
lected rhodopsin-based models identified four novel H4 li-
gands including compound 3 containing a novel thiocarba-
mide scaffold providing an interesting starting point for fur-
ther medicinal chemistry optimization.
Computational Methods and Experimental
Retrospective Screening. H4 reference agonists and antago-
nists were collected from the Thomson Reuters Integrity
database (Thomson Reuters, integrity.thomson-pharma.-
com) and decoys were selected from the NCI database (Na-
tional Cancer Institute, version: January 2010), which con-
tains considerably diverse compounds.
Docking was carried out by FlexX (version 3.1, BioSolveIT
GmbH) with default settings, except: (i) ring conformers
were generated by built-in Corina, (ii) pseudo R/S and E/Z
configurations were allowed to be inverted and (iii) input
protonation states of the ligands were preserved.
HH4R models were derived from our previous study. We
extracted one structure after every ns simulation yielding
21–21 snapshots of hH4R (including the initial structures).
The per-residue RMSD of the heavy atoms of interacting
sidechains were calculated in order to show the structural
difference between frames selected (See Supporting Infor-
mation).
We have analyzed the predictive power of all these
models by enrichment tests, where enrichment factors (de-
fined as the percentage of actives retrieved divided by the
percentage fraction of the database considered) were cal-
culated.
Prospective Screening. After the most promising models
have been selected, a smaller subset of the in-house com-
pound database of Gedeon Richter (43394 compounds)
was prefiltered for prospective screening. H4 affinity of the
best scoring hits has been measured by H4 radioligand
binding assay. Purity of in vitro hits have been evaluated
and compounds having a purity above 90 % were reported
as confirmed hits.
 2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Detailed description of the calculation protocols, radioli-
gand binding assay and purity evaluation is available as
Supporting Information.
Acknowledgements
We kindly acknowledge BioSolveIT for providing license for
its docking tool FlexX for this study. The authors also thank
Mrk Sndor, Mrton Vass and Gyçrgy T. Balogh for their
help.
References
[1] R. Kiss, G. M. Keserű, in Virtual Screening: Principles, Challenges,
and Practical Guidelines (Ed: C. Sotriffer), Wiley-VCH, Weinheim,
2011, pp. 381 – 410.
[2] R. Kiss, B. Kiss, A. Kçnczçl, F. Szalai, I. Jelinek, V. Lszl, B.
Noszl, A. Falus, G. M. Keserű, J. Med. Chem. 2008, 51, 3145 –
3153.
[3] R. Kiss, T. Polgr, A. Kirabo, J. Sayyah, N. C. Figueroa, A. F. List,
L. Sokol, K. S. Zuckerman, M. Gali, K. S. Bisht, P. P. Sayeski, G. M.
Keserű, Bioorg. Med. Chem. Lett. 2009, 19, 3598 – 3601.
[4] P. Khare, A. K. Gupta, P. K. Gajula, K. Y. Sunkari, A. K. Jaiswal, S.
Das, P. Bajpai, T. K. Chakraborty, A. Dube, A. K. Saxena, J. Chem.
Inf. Model. 2012, 52, 777 – 791.
[5] S. L. Li, Y. Zhou, W. Q. Lu, Y. Zhong, W. L. Song, K. D. Liu, J.
Huang, Z. J. Zhao, Y. F. Xu, X. F. Liu, H. L. Li, J. Chem. Inf. Model.
2011, 51, 2939 – 2947.
[6] H. A. Carlson, J. A. McCammon, Mol. Pharmacol. 2000, 57,
213 – 218.
[7] J. H. Lin, Curr. Top Med. Chem. 2011, 11, 171 – 178.
[8] O. Korb, T. S. G. Olsson, S. J. Bowden, R. J. Hall, M. L. Verdonk,
J. W. Liebeschuetz, J. C. Cole, J. Chem. Inf. Model. 2012, 52,
1262 – 1274.
[9] S. Y. Huang, X. Q. Zou, Proteins 2007, 66, 399 – 421.
[10] C. N. Cavasotto, J. A. Kovacs, R. A. Abagyan, J. Am. Chem. Soc.
2005, 127, 9632 – 9640.
[11] B. Jojrt, R. Kiss, B. Viskolcz, G. M. Keserű, J. Chem. Inf. Model.
2008, 48, 1199 – 1210.
[12] A. Tarcsay, G. Paragi, M. Vass, B. Jjrt, F. Bogr, G. M. Keserű, J.
Chem. Inf. Model. 2013, 53, 2990 – 2999.
[13] E. A. Gattica, C. N. Cavasotto, J. Chem. Inf. Model. 2012, 52, 1 –
6.
[14] A. L. Hopkins, C. R. Groom, A. Alex, Drug. Discov. Today 2004,
9, 430 – 431.
[15] G. M. Keserű, G. M. Makara, Nat. Rev. Drug Discov. 2009, 8,
203 – 212.
[16] Y. Tanrikulu, E. Proschak, T. Werner, T. Geppert, N. Todoroff, A.
Klenner, T. Kottke, K. Sander, E. Schneider, R. Seifert, H. Stark, T.
Clark, G. Schneider, ChemMedChem 2009, 4, 820 – 827.
[17] ChEMBL database – accessed on the 24th of November 2013.
Received: April 19, 2013
Accepted: February 6, 2014
Published online: March 20, 2014
Mol. Inf. 2014, 33, 264 – 268 268