вход по аккаунту


Ligand-Supported Homology Modeling of G-Protein-Coupled Receptor Sites Models Sufficient for Successful Virtual Screening.

код для вставкиСкачать
Protein Models
Ligand-Supported Homology Modeling of
G-Protein-Coupled Receptor Sites: Models
Sufficient for Successful Virtual Screening**
Andreas Evers and Gerhard Klebe*
Recent advances in the various genome-sequencing projects
have opened the floodgates to thousands of protein sequences
that are possibly new targets for drug discovery.[1] Accordingly, computer techniques are required for the reliable
prediction of protein structures accurate enough to serve as a
platform for structure-based drug design. We have developed
a new approach that models proteins by homology. Because
the bound ligand molecules serve as restraints, more relevant
geometries of protein binding sites result.[2] Initial homology
models of the target protein are optimized iteratively by
including information about bioactive ligands as spatial
restraints (Figure 1).
Figure 1. Schematic overview of the concept followed in our MOBILE
approach. Starting with the structure of one or more template proteins, we generate several preliminary homology models of our target
protein using MODELLER.[3] Next, one or more ligands known to bind
to the target protein are placed into an averaged binding-site representation of the generated ensemble of superimposed models.[4] Improved
protein models are then generated with MODELLER, now explicitly
considering the docked ligand(s) as additional restraints. The ligands
are considered in terms of knowledge-based pair potentials extracted
from the protein–ligand scoring function DrugScore.[5] After the resulting generated complexes are scored with DrugScore, a final model is
chosen from the individual models that best explains the observed
ligand binding. This complex can be optimized further by selecting
and combining portions from different models. Finally, the model is
refined by optimizing the side-chain-to-ligand interactions using a
common force field.[6]
[*] A. Evers, Prof. Dr. G. Klebe
Institut f6r Pharmazeutische Chemie der Universit9t Marburg
Marbacher Weg 6, 35032 Marburg (Germany)
Fax: (+ 49) 06421-28-28994
[**] The authors are grateful to Dr. Arne Svensson (AstraZeneca,
MClndal) for determining the binding affinities of the seven test
compounds described in this report.
2004 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Our approach, MOBILE (Modeling Binding Sites Including Ligand Information Explicitly), has been parameterized
and validated by reproducing binding-site models of proteins
for which the structure is actually known. Subsequently, we
applied this method to the discovery of neurokinin-1 (NK1)
receptor antagonists. A ligand with binding affinity in the
submicromolar range was found for this protein, which
belongs to the family of G-protein-coupled receptors
In structure-based drug design, knowledge of the threedimensional structure of a target protein is of utmost
importance. It is either determined by X-ray crystallography
or by NMR spectroscopy. However, the rate at which protein
sequences are presently being discovered exceeds by far the
rate at which protein structures can be determined experimentally. Thus, for a considerable number of putative drug
targets, the three-dimensional structure will not be readily
available. In such cases the most reliable computer-based
technique for generating a three-dimensional protein structure is homology modeling.[7] However, homology modeling
only considers information available from the related protein
structures. The model remains at a rather approximate level if
in the target protein several amino acids of the active site are
replaced with respect to those in the template protein(s).
Usually, in a drug-design project, binding data about ligands
become available, for example, from high-throughput screening, before the spatial structure is elucidated. In due course,
ligand-based structure–activity relationships are established
(by 3D QSAR methods)[8] to pinpoint features responsible for
trends in the binding affinity. These methods generate some
kind of a generalized blueprint of the putative binding pocket
of the target protein without giving explicit information about
its actual composition. These methods further support ligand
optimization; however, they are limited with respect to the
design of novel or alternative molecular skeletons. Our new
method tries to combine the advantages of both homology
modeling and ligand-based 3D QSAR analysis.
In a real-life test scenario we applied MOBILE to the
neurokinin-1 (NK1) receptor belonging to the superfamily of
GPCRs mediating responses to, for example, visual, olfactory,
hormonal, and neurotransmitter signals. GPCRs are one of
the most relevant target families in small-molecule drug
design. Currently, 50 % of all marketed drugs address
GPCRs.[9] Since GPCRs are membrane-bound proteins,
their expression, purification, crystallization, and structure
determination remain difficult. So far, only the structure of
bovine rhodopsin could be determined with sufficiently high
resolution.[10] Its crystal structure serves as a template
reference for homology modeling of all other members of
the GPCR family. Because of the limited accuracy, drug
discovery based on virtual screening with rhodopsin-based
GPCR models has not yet been reported in literature.
Instead, successful computer-aided drug discovery for
GPCRs was achieved by applying ligand-based virtual screening techniques.[11]
A first step towards drug discovery with rhodopsin-based
protein models was attempted by Bissantz et al.,[12] who
recently demonstrated that their homology models of the
dopamine D3, muscarinic M1, and vasopressin V1a receptors
DOI: 10.1002/anie.200352776
Angew. Chem. Int. Ed. 2004, 43, 248 –251
were reliable enough to retrieve known antagonists by virtual
screening from a database, which also comprised randomly
selected drug-like molecules. Another GPCR modeling
technique that does not rely on the crystal structure of
bovine rhodopsin is PREDICT.[13] In screening for novel
binders based on their GPCR models, the authors report hit
rates of 10 to 24 %.
The NK1 receptor belongs to the series of tachykininbinding receptors (NK1, NK2, and NK3). They selectively
bind the peptide neurotransmitters substance P, neurokinin A, and neurokinin B, respectively. Substance P plays a
role in the transmission of pain and is involved in inflammation and immune response. The probably best studied NK1
antagonist is CP-96 345. Information on its bioactive conformation was obtained by several theoretical and experimental studies.[14] Through mutational studies[15] and comparative affinity determinations based on CP-96 345 binding,
the essential amino acids involved in ligand recognition could
be identified and translated into a crude topographical
interaction model (see Figure 3 b). According to our homology model, the binding pocket of CP-96 345 overlaps with the
binding site of retinal in bovine rhodopsin.
The sequence identity between bovine rhodopsin and the
NK1 receptor amounts to 21 % (Figure 2). If only the
transmembrane regions are considered, this increases to
27 %. No sequence similarity can be detected to the retinal
binding-site of bovine rhodopsin, which renders the homology
modeling particularly challenging. With our approach, 100
preliminary homology models based on the crystal structure
of rhodopsin were generated. Next, CP-96 345 was docked
rigidly into the modeled binding pockets with AutoDock
using the ligand conformation observed in its crystal structure
(CSD reference code LEWCUL; Figure 3 a). From the set of
docked ligand conformations, we selected the solution that
best reproduced the key interactions depicted in the topographical model (Figure 3 b). Subsequently, new homology
models were generated explicitly considering bound CP96 345. The best-scored amino acids were merged in a
combinatorial fashion, and finally a model without unfavorable van der Waals interactions was obtained.
Figure 3 c shows the ligand-binding site of the resulting
complex, which convincingly reproduces the proposed interactions in the 2D interaction model (Figure 3 b).[14] The
quality of the model was validated by probing its ability to
accommodate other known NK1 antagonists from structurally diverse compound classes. Common chemical features of
known NK1 antagonists and published mutation data identifying the key interaction points of the ligands with the NK1
receptor were used to establish a pharmacophore hypothesis
(see Figure 3 d). This model was used to retrieve ligands from
seven databases containing a total of over 800 000 compounds.
A search strategy similar to that reported previously by us
for the search of enzyme inhibitors[17] was applied. We focused
only on “leadlike” compounds with up to seven rotatable
bonds and a molecular weight of less than 450 Da. A 2D filter
was applied to extract compounds containing at least two
phenyl rings and one hydrogen-bond acceptor. The program
UNITY was then used to constrain the mutual spatial
arrangement of the aromatic rings and the hydrogen-bond
acceptor (Figure 3 b and d), considering the protein environment in terms of excluded volume constraints. The remaining
11 109 compounds were flexibly docked into the modeled
NK1 binding pocket using the program FlexX-Pharm.[18] As a
reference, the pharmacophore depicted in Figure 3 d was used
Figure 2. Sequence alignment of bovine rhodopsin and the NK1 receptor. All residues known from mutational studies to be involved in binding CP96 345 are marked in gray. These residues are Gln 165 (according to the Ballesteros–Weinstein numbering scheme[16] 4.60), Glu 193 (5.35), His 197
(5.39), Ile 204 (5.46), His 265 (6.52), and Tyr 272 (6.59).
Angew. Chem. Int. Ed. 2004, 43, 248 –251
2004 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
protein and accommodated ligand. In
particular, highly flexible ligands and
ligands having unsatisfied H-bond
donors or acceptors and unoccupied
voids along the interface were eliminated.
Finally, seven compounds (Scheme 1)
were selected for in vitro testing. We
chose a radioligand binding assay on
whole CHO (Chinese Hamster Ovary)
cells (with substance P as the radioligand), which is sensitive enough to
detect ligands with a binding affinity of
at least 1 mm. Of the seven compounds,
one showed affinity (ASN-1 377 642 from
AEPC, see Figure 4) in the submicromolar range, exhibiting a Ki value of 251 nm.
ASN-1 377 642 meets the criterion
defined in our screening query.
We can conclude that our ligandsupported homology modeling produces
binding-site models that successfully
serve as a platforms for structure-based
drug design. A major advantage is the fact
Figure 3. The NK1 receptor and its antagonist CP-96 345. a) Top view (extracellular) of the NK1
that information about known ligands is
receptor homology model complexed with CP-96 345. b) Schematic representation of postunot only included into the protein-modlated interactions of the NK1 receptor with CP-96 345, established by analyzing the published
eling process but also considered in the
mutational data and affinity measurements of NK1 antagonists. The arrows indicate proposed
key interactions between the receptor and the ligand. The most important interactions essensubsequent screening: Structural features
tial for the binding of NK1 antagonists are 1) the hydrogen bond between Gln 165(4.60)-NH
about known NK1 antagonists were used
and a corresponding acceptor of the antagonist and 2) an amine–arene interaction established
to establish the pharmacophoric elements
between His 197 (5.39) and the the aromatic moiety B. Analysis of NK1 antagonists from structhat drive our hierarchical 2D and 3D
turally diverse classes reveals that the aromatic group C is also required for binding. c) Modscreening strategy. In this way the protein
eled complex of the NK1 receptor with CP-96 345. The dashed lines indicate the key interacmodel restrained the 3D screening protions proposed in the topographical interaction model depicted in part b. d) Structure-based
cess and guided the generation of ligand
pharmacophore hypothesis based upon data from known NK1 antagonists. The hypothesis is
characterized by three major features: 1) a hydrogen bond between the Gln 165(4.60)-NH and
conformations in the binding pocket.
a corresponding acceptor (here, N2) of the ligand; 2) the aromatic moiety B which interacts
Finally, these docking solutions were
with His 197 (5.39) in amine–arene interactions; and 3) a further aromatic group C.
ranked by evaluating the interactions
between the protein and the ligand. Considering that CP-96 345 and ASN1 377 642 have similar pharmacophore features, but distinctly
as an additional constraint within the docking process.
different molecular frameworks, we believe that our hit would
Subsequently, binding affinities were estimated using Drugnot have been one of the top hits in a solely ligand-based
Score. After minimizing the produced complexes of the bestscreening approach.
ranked solutions using MOLOC,[6] we inspected the best hits
It should be noted that the success of this method depends
visually and analyzed their adopted binding modes, the
on the sequence identity between the target and template
agreement with the proposed 2D interaction model (Figprotein in the ligand binding pocket. If this agreement is low,
ure 3 b), and the mutual surface complementarity between
Scheme 1. Compounds that were tested for antagonism of the NK1 receptor. ASN-1 377 642 has a Ki value of 251 nm.
2004 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Angew. Chem. Int. Ed. 2004, 43, 248 –251
Figure 4. Modeled binding mode of ASN-1 377 642, which was discovered as a potent NK1 antagonist by virtual screening based on our
homology model. The pharmacophore hypothesis is convincingly well
matched. The essential hydrogen bond to Gln 165(4.60)-NH is established through the carbonyl oxygen atom of an amide group.
Baker, J. Med. Chem. 1995, 38, 4793 – 4805; c) G. J. Boks, J. P.
Tollenaere, J. Kroon, Bioorg. Med. Chem. 1997, 5, 535 – 547.
a) T. M. Fong, M. A. Cascieri, H. Yu, A. Bansal, C. Swain,
Nature 1993, 362, 350 – 353; b) T. M. Fong, H. Yu, M. A. Cascieri,
D. Underwood, C. Swain, C. D. Strader, J. Biol. Chem. 1994, 269,
14 957 – 14 961; c) T. M. Fong, H. Yu, M. A. Cascieri, D. Underwood, C. Swain, C. D. Strader, J. Biol. Chem. 1994, 269, 2728 –
J. A. Ballesteros, H. Weinstein, Methods Neurosci. 1995, 25, 366 –
a) S. Gruneberg, B. Wendt, G. Klebe, Angew. Chem. 2001, 113,
404 – 408; Angew. Chem. Int. Ed. 2001, 40, 389 – 393; b) S.
Gruneberg, M. T. Stubbs, G. Klebe, J. Med. Chem. 2002, 45,
3588 – 3602; c) R. Brenk, L. Naerum, U. Gradler, H. D. Gerber,
G. A. Garcia, K. Reuter, M. T. Stubbs, G. Klebe, J. Med. Chem.
2003, 46, 1133 – 1143.
S. A. Hindle, M. Rarey, C. Buning, T. Lengauer, J. Comput.Aided Mol. Des. 2002, 16, 129 – 149.
additional information is required, for example, from mutational data, and from a high-affinity ligand conformation that
is close to the bioactive one.
Received: September 3, 2003 [Z52776]
Published Online: December 2, 2003
Keywords: computer chemistry · G proteins · protein models ·
[1] a) A. L. Hopkins, C. R. Groom, Nat. Rev. Drug Discovery 2002,
1, 727 – 730; b) J. Drews, Science 2000, 287, 1960 – 1964.
[2] A. Evers, H. Gohlke, G. Klebe, J. Mol. Biol. 2003, 334, 327 – 345.
[3] A. Sali, T. L. Blundell, J. Mol. Biol. 1993, 234, 779 – 815.
[4] a) A. Schafferhans, G. Klebe, J. Mol. Biol. 2001, 307, 407 – 427;
b) F. Osterberg, G. M. Morris, M. F. Sanner, A. J. Olson, D. S.
Goodsell, Proteins Struct. Funct. Genet. 2002, 46, 34 – 40.
[5] H. Gohlke, M. Hendlich, G. Klebe, J. Mol. Biol. 2000, 295, 337 –
[6] P. R. Gerber, J. Comput.-Aided Mol. Des. 1998, 12, 37 – 51.
[7] D. Baker, A. Sali, Science 2001, 294, 93 – 96.
[8] a) 3D QSAR in Drug Design. Ligand–Protein Complexes and
Molecular Similarity. (Eds.: H. Kubinyi, G. Folkers, Y. C.
Martin), Kluwer/ESCOM, Leiden, 1998; b) 3D QSAR in Drug
Design. Recent Advances. (Eds.: H. Kubinyi, G. Folkers, Y. C.
Martin), Kluwer/ESCOM, Leiden, 1997.
[9] T. Klabunde, G. Hessler, ChemBioChem 2002, 3, 928 – 944.
[10] a) D. C. Teller, T. Okada, C. A. Behnke, K. Palczewski, R. E.
Stenkamp, Biochemistry 2001, 40, 7761 – 7772; b) K. Palczewski,
T. Kumasaka, T. Hori, C. A. Behnke, H. Motoshima, B. A. Fox, I.
Le Trong, D. C. Teller, T. Okada, R. E. Stenkamp, M. Yamamoto, M. Miyano, Science 2000, 289, 739 – 745.
[11] a) S. Flohr, M. Kurz, E. Kostenis, A. Brkovich, A. Fournier, T.
Klabunde, J. Med. Chem. 2002, 45, 1799 – 1805; b) D. P. Marriott,
I. G. Dougall, P. Meghani, Y. J. Liu, D. R. Flower, J. Med. Chem.
1999, 42, 3210 – 3216.
[12] C. Bissantz, P. Bernard, M. Hibert, D. Rognan, Proteins Struct.
Funct. Genet. 2003, 50, 5 – 25.
[13] O. M. Becker, S. Shacham, Y. Marantz, S. Noiman, Curr. Opin.
Drug Discovery Dev. 2003, 6, 353 – 361.
[14] a) Y. Takeuchi, E. F. B. Shands, D. D. Beusen, G. R. Marshall, J.
Med. Chem. 1998, 41, 3609 – 3623; b) C. J. Swain, E. M. Seward,
M. A. Cascieri, T. M. Fong, R. Herbert, D. E. MacIntyre, K. J.
Merchant, S. N. Owen, A. P. Owens, V. Sabin, M. Teall, M. B.
VanNiel, B. J. Williams, S. Sadowski, C. Strader, R. G. Ball, R.
Angew. Chem. Int. Ed. 2004, 43, 248 –251
2004 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Без категории
Размер файла
189 Кб
site, homology, sufficient, modeling, supported, ligand, virtual, model, screening, protein, successful, receptov, coupled
Пожаловаться на содержимое документа