Structural requirements for factor Xa inhibition by 3-oxybenzamides with neutral P1 substituents: Combining X-ray crystallography, 3D-QSAR, and tailored scoring functions

Article abstract of DOI:10.1021/jm049187l

The design, synthesis, and structure-activity relationship of 3-oxybenzamides as potent inhibitors of the coagulation protease factor Xa are described on the basis of X-ray structures, privileged structure motifs, and SAR information. A total of six X-ray structures of fXa/inhibitor complexes led us to identify the major protein-ligand interactions. The binding mode is characterized by a lipophilic dichlorophenyl substituent interacting with Tyr228 in the protease S1 pocket, while polar parts are accommodated in S4. This alignment in combination with docking allowed derivation of 3D-QSAR models and tailored scoring functions to rationalize biological affinity and provide guidelines for optimization. The resulting models showed good correlation coefficients and predictions of external test sets. Furthermore, they correspond to binding site topologies in terms of steric, electrostatic, and hydrophobic complementarity. Two approaches to derive tailored scoring functions combining binding site and ligand information led to predictive models with acceptable predictions of the external set. Good correlations to experimental affinities were obtained for both AFMoC (adaptation of fields for molecular comparison) and the novel TScore function. The SAR information from 3D-QSAR and tailored scoring functions agrees with all experimental data and provides guidelines and reasonable activity estimations for novel fXa inhibitors.

Full text of DOI:10.1021/jm049187l

3290
J. Med. Chem. 2005, 48, 3290-3312
Structural Requirements for Factor Xa Inhibition by 3-Oxybenzamides with
Neutral P1 Substituents: Combining X-ray Crystallography, 3D-QSAR, and
Tailored Scoring Functions
Hans Matter,* David W. Will, Marc Nazare´, Herman Schreuder, Volker Laux, and Volkmar Wehner
DI&A Chemistry, Aventis Pharma Deutschland GmbH, A Company of the Sanofi-Aventis Group, Building G 878,
D-65926 Frankfurt am Main, Germany
Received October 12, 2004
The design, synthesis, and structure-activity relationship of 3-oxybenzamides as potent
inhibitors of the coagulation protease factor Xa are described on the basis of X-ray structures,
privileged structure motifs, and SAR information. A total of six X-ray structures of fXa/inhibitor
complexes led us to identify the major protein-ligand interactions. The binding mode is
characterized by a lipophilic dichlorophenyl substituent interacting with Tyr228 in the protease
S1 pocket, while polar parts are accommodated in S4. This alignment in combination with
docking allowed derivation of 3D-QSAR models and tailored scoring functions to rationalize
biological affinity and provide guidelines for optimization. The resulting models showed good
correlation coefficients and predictions of external test sets. Furthermore, they correspond to
binding site topologies in terms of steric, electrostatic, and hydrophobic complementarity. Two
approaches to derive tailored scoring functions combining binding site and ligand information
led to predictive models with acceptable predictions of the external set. Good correlations to
experimental affinities were obtained for both AFMoC (adaptation of fields for molecular
comparison) and the novel TScore function. The SAR information from 3D-QSAR and tailored
scoring functions agrees with all experimental data and provides guidelines and reasonable
activity estimations for novel fXa inhibitors.
1. Introduction
reviews12-17 and other publications18,19 demonstrate the
high interest of the pharmaceutical industry in the
development of novel fXa inhibitors.
There is tremendous interest in the development of
new, orally active anticoagulants for the treatment and
prevention of thrombotic diseases, with therapeutic
advantages over pharmaceuticals such as heparin and
warfarin. Thrombotic diseases, such as deep vein throm-
bosis and stroke, are major causes of mortality in
Europe and the U.S.1 The blood coagulation serine
protease factor Xa (fXa) is an attractive target because
it is the central enzyme in the activation cascade of the
coagulation system,2 linking the intrinsic and extrinsic
pathways to the common coagulation pathway.3,4 In
combination with factor Va, fXa activates prothrombin
on a phospholipid surface to generate the coagulation
protease thrombin.5 It is not known to be involved in
processes other than hemostasis. Thrombin then con-
verts fibrinogen to fibrin, inducing clot formation and
platelet aggregation,6 which can both be related to
serious pathological situations. The inhibition of fXa
compared to thrombin might allow the effective control
of thrombogenesis with a minimal effect upon bleed-
ing7-11 because fXa inhibitors should affect coagulation
specifically. Furthermore, inhibition of fXa is seen to
be more efficacious because one fXa generates multiple
thrombin molecules. Inhibition of fXa should prevent
production of new thrombin without affecting the basal
thrombin level necessary for primary hemostasis. Hence,
it is anticipated that inhibition of factor Xa should
prevent thrombus formation without compromising
normal hemostasis and platelet function. Many recent
The mature form of factor X with a 139-residue light
chain and a 303-residue heavy chain is synthesized in
the liver and secreted after post-translational modifica-
tions into the blood as zymogen. It is activated by the
factor VIIa/tissue factor complex in the extrinsic path-
way, initiated by vascular damage or by the factor IXa/
factor VIIIa complex in the intrinsic pathway. The fXa
heavy chain contains a serine protease domain in a
trypsin-like closed â-barrel fold encompassing the active
triad Ser195-His57-Asp102 and two essential protein
subsites S1 and S4, which are often explored in structure-
guided drug design.20
In this publication we report the design and structure-
activity relationship of a series of nonchiral 3-oxybenz-
amides (synthesis shown in Scheme 1) as inhibitors of
fXa21 by means of X-ray crystallography, 3D-QSAR, and
tailored scoring functions. These inhibitors contain
neither a benzamidine nor a guanidine moiety but
employ a neutral substituent for binding into the
protease S1 pocket. First, the synthesis of a targeted
compound library focused toward fXa led to the identi-
fication of a guanidine containing hit structure 57 (Table
1) bearing a benzoic acid scaffold and a neutral dichlo-
rophenyl substituent. Although it was expected that this
guanidine would bind to the fXa S1 pocket, X-ray
structure determination showed that surprisingly the
neutral dichlorophenyl substituent interacted with S1
and the guanidine with S4. This protein-ligand inter-
action motif was systematically explored for subsequent
* To whom correspondence should be addressed. Phone: ++49-69-
305-84329. Fax: ++49-69-331399. E-mail: hans.matter@aventis.com.
10.1021/jm049187l CCC: $30.25 © 2005 American Chemical Society
Published on Web 04/13/2005

Factor Xa Inhibition
Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9 3291
Scheme 1. Synthesis of 3-Oxybenzamidesa
a Reagents and conditions: (a) MeOH, HCl; (b) 2-(2,4-dichlorophenyl)ethanol, DEAD, PPh3-polystyrene, THF/room temp, 16 h; (c) (i)
NaOH (aq), dioxane/60 °C, 1 h, (ii) HCl/water, precipitation; (d) amine, TOTU, N-ethylmorpholine, DMF.
design using high-resolution fXa X-ray structures of
derivatives, flexible docking, tailored scoring functions,
and 3D-QSAR analysis. Our design rationale was based
on X-ray structures of fXa in uncomplexed form22 and
with inhibitors19,23-25 plus knowledge of privileged
motifs directed toward the S1 pocket and structure-
activity information on privileged substructures ac-
cumulated in the project. Comparative molecular field
analysis (CoMFA)26-28 and comparative molecular simi-
larity index analysis (CoMSIA)29 are used to correlate
molecular property fields to biological activities based
on the X-ray structures of some analogues providing the
active inhibitor conformation for alignment. The super-
position of all other molecules onto these templates
produced consistent models in agreement with binding
site requirements. The contour maps from 3D-QSAR
models enhance the understanding of electrostatic,
hydrophobic, and steric requirements for ligand binding,
guiding the design of inhibitors to regions where struc-
tural variations reveal a correlation to biological proper-
ties.
Tailored scoring functions were also derived on the
basis of two approaches to establish predictive models
for structure-based optimization. The difference to 3D-
QSAR is the incorporation of binding site information
not only for alignment but also to derive descriptors for
protein-ligand interactions. First, the recently intro-
duced AFMoC (adaptation of fields for molecular com-
parison) approach was applied.30 Knowledge-based pair
potentials are adapted to a binding site by considering
ligand information. Atom-type specific interaction fields
capturing binding site characteristics and ligand com-
plementarity are correlated to biological affinities. In
contrast, the in-house approach TScore31 captures pro-
tein-ligand interaction on an amino acid level. For each
binding site residue and ligand, terms describing hy-
drogen-bonding, lipophilicity, and steric contacts pro-
duce a protein-ligand interaction profile. Their statis-
tical analysis by correlating them to affinities led to
relevant models. Hence, both approaches led to tailored
scoring functions with predictive power and interpret-
ability in agreement with X-ray structures and 3D-
QSAR. Some promising inhibitors of this protease
emerged from this study as starting points for further
optimization.
153 nM.21 This compound was expected to bind with its
guanidine in the S1 pocket. However, its X-ray crystal
structure in complex with factor Xa (see below) revealed
that the neutral dichlorophenyl substituent binds to the
S1 pocket, while the guanidine interacts with S4. This
“chloro-binding mode”, which was independently dis-
covered by others for thrombin32 and fXa,23,24 was
explored for further design. Although cationic interac-
tions of ligand substructures with amino acids in both
the S1 and S4 subsite are favorable for in vitro affinity,
permanently charged groups might be detrimental for
oral bioavailability. Hence, the replacement of basic
moieties in S1 by neutral substituents and the reduction
of the basic nature of the S4 directed substructure were
important considerations guiding our structure-based
design efforts. We were able with this achiral motif to
orient different substituents toward essential fXa sub-
pockets, namely, S1, S4, and the so-called “ester binding
pocket” (EBP), which is located adjacent to S1.21 The
arginine in 57 was replaced by synthesis of a library of
330 amides selected for their fit into S4 by docking,
resulting in a series of potent, less basic pyridines,
pyrimidines, and piperidines interacting with this sub-
site.21 The 4-piperidylpyridine motif resulted in very
potent fXa inhibitors, for example, 50 with a Ki of 18
nM. This compound then served as a lead for all further
evaluations.
For selected analogues high-resolution fXa X-ray
structures were obtained, thus confirming individual
structure-based design iterations, further underscoring
our initial assumption of a consistent binding mode for
this series. Critical determinants for ligand binding to
fXa were deduced from this combination of structure-
based design techniques and 3D-QSAR. The 3D-QSAR
models then allowed affinity predictions complementing
structure-based scoring functions for evaluation of
synthesis candidates. Their interpretation uncovered
important steric, electrostatic, and hydrophobic fea-
tures, which are linked to fXa affinity. On the basis of
SAR, docking, and X-ray crystallography, novel com-
pounds were synthesized. The final set of 80 molecules
as training set for this study and 27 test molecules are
summarized in Table 1, where the training set is
indicated as SAR set 1 and the test set is indicated as
SAR set 2.
3. Methods
2. Design of Factor Xa Inhibitors
3.1. Chemistry and Enzyme Assay. Synthesis of all
compounds in Table 1 was achieved from commercially avail-
able or easily accessible 3-hydroxy-4-substituted-benzoic acid
methyl esters in a conventional synthesis sequence, as de-
scribed earlier.21,33 Esterification in methanol led to the central
The 3-oxybenzamide scaffold resulted in potent and
selective fXa inhibitors after screening of targeted
compound libraries. This led to the identification of the
arginine containing hit 57 (Table 1) with a Ki value of

3292 Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9
Table 1. Chemical Structure and Activities for 3-Oxybenzamides
Matter et al.

Factor Xa Inhibition
Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9 3293
Table 1. (Continued)

3294 Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9
Table 1. (Continued)
Matter et al.

Factor Xa Inhibition
Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9 3295
Table 1. (Continued)

3296 Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9
Table 1. (Continued)
Matter et al.
a Corresponding numbering of compounds in ref 21.
Table 2. Data Collection and Refinement Statistics for Six Factor Xa X-ray Structures
inhibitor
50
63
57
87
56
37
detector
space group
a, Å
b, Å
c, Å
obsd reflns
unique reflns
resoln, Å
R-sym, %
marccd
P212121
57.0
72.3
78.1
33 939
9120
2.7
marccd
P212121
56.0
71.7
78.6
21 647
5987
3.1
12.8
98.2
2248
32
mar300
P212121
56.3
71.7
77.5
57 893
18 769
2.1
mar300
P212121
56.07
71.53
78.36
21 857
9132
2.65
mar300
P212121
56.7
71.96
78.55
19 540
5126
3.3
mar300
P212121
56.34
71.95
78.07
43 464
11 393
2.5
8.9
6.3
10.8
14.98
99.4
6.6
completeness, %
protein atoms
inhibitor atoms
calcium ions
water molecules
R-factor, %
98.0
2249
35
1
218
19.3
27.7
0.009
1.33
24.8
0.86
99.2
2249
37
1
288
22.0
29.6
0.009
1.60
24.8
2.11
95.1
99.3
2249
33
2249
35
2240
36
1
294
1
1
1
284
283
241
15.7
23.0
0.015
1.69
25.4
0.81
17.2
15.4
19.8
R-free, %
24.6
27.7
24.8
rmsd bond length, Å
rmsd bond angle, deg
rmsd dihedrals, deg
rmsd impropers, deg
0.008
1.34
0.012
1.35
0.013
1.29
24.8
0.76
24.5
0.81
24.5
0.75
precursor for the attachment of the arylethylene moiety
directed toward S1. This etherification was performed under
Mitsunobu conditions.34 Attempted alkylation of the 3-hydroxy-
4-substituted-benzoic methyl esters via tosylates or mesylates
of the arylethylene alcohols under various conditions was less
successful because of extensive elimination of the activated
arylethylene moiety. After saponification of the methyl ester,
the acid was subjected to the final amide coupling using TOTU
(O-((ethoxycarbonyl)cyanomethyleneamino)-N,N,N′,N′-tetra-
methyluronium tetrafluoroborate) as activating agent. The
biological assay was reported earlier.35 Enzyme inhibition (pKi)
is expressed as log[(1/K )(1000)].
refined structures of fXa complexes solved previously with the
same crystal packing as the current complexes. The bound
inhibitors were omitted from the search models. Energy-
restrained least-squares refinement was carried out using
X-PLOR.37 This refinement was started with rigid body refine-
ment to adjust for small differences in cell dimensions, followed
by energy minimization and individual temperature factor
refinement. At this stage the 2Fo - Fc and Fo - Fc maps were
inspected and the inhibitors were fitted. Solvent molecules
were included if they were on sites of difference electron
density with values above 3.5σ and if they were within 3.5 Å
of the protein molecule or a water molecule. After two to three
additional rounds of manual inspection, rebuilding, and refine-
ment, final models were obtained with R factors between 15.4%
and 22.0% and free R factors between 23% and 29.6% plus
good geometry. The EGF-1 domain is not visible in the electron
density maps probably because of disorder, and the rather high
free R factors might relate to this disordered EGF-1 domain.
The statistics of the crystallographic refinement are listed in
Table 2.
3.2. X-ray Structuire Analysis. 3.2.1. Crystallization.
Purified human fXa was purchased from Enzyme Research
Lab (South Bend, IN). The Gla domain was removed, and the
Gla-less factor Xa was crystallized in hanging drops as
described earlier.23
3.2.2. Data Collection and Processing. Crystals were
soaked in a 5 µL reservoir solution containing ∼20 mM or
saturated inhibitor (depending on the solubility) for 24-72 h.
Data were collected at cryotemperatures. The crystals were
picked up with a fiber loop, soaked for a few seconds in a
solution containing 20% glycerol and 5-20 mM inhibitor in
reservoir solution, and flash-frozen in a stream of gaseous
nitrogen at 100 K. The X-ray intensity data were collected on
a 130 mm Mar CCD detector mounted on an Elliot GX-13 “big
wheel” rotating anode generator (Nonius, The Netherlands)
and operated at 40 kV and 55 mA or on a Mar300 imaging
plate (X-ray research, Germany) mounted on a FR591 rotating
anode (Nonius, The Netherlands) and operated at 40 kV and
80 mA. Data processing and scaling were carried out using
XDS.36 Data collection and refinement statistics for selected
crystal structures are presented in Table 2.
3.3. Computational Procedures.
3.3.1. Docking Studies. FXa crystal structures from
RSCB38 and our database were used for docking. After analysis
of protein-ligand interactions using the program GRID,39
molecules were oriented and minimized within the binding site
or automatically docked in different orientations using the
program QXP40 with a modified version of the AMBER force
field.41 Selected protein side chains were treated as flexible
after comparative analysis of fXa X-ray structures. A structur-
ally conserved water situated in the S4 pocket was included
in energy minimizations. The alignment for all molecules from
Table 1 after visual inspection of QXP docking modes served
as the basis for 3D-QSAR studies and tailored scoring func-
tions. To further assess the influence of the force field on the
model quality, all compounds were optimized using the
MMFF94s force field42 in SYBYL.43 This alignment was also
3.2.3. Structure Solution and Crystallographic Re-
finement. The structures were solved by molecular replace-
ment. The search models were made from the coordinates of

Factor Xa Inhibition
Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9 3297
used to produce statistical models. Protein/ligand complexes
were minimized using quasi-Newton-Raphson (BFGS) or
conjugate gradient (CG) procedures with all protein atoms
being rigid. Introducing flexibility into amino acids within 4
Å around the ligands did not improve the statistical results.
The program MOLCAD44 was used to visualize properties such
as lipophilicity45,46 and cavity depth on solvent accessible
protein surfaces.47
3.3.2. 3D-QSAR. Default settings were used for CoMFA and
CoMSIA, if not otherwise indicated. For CoMFA, steric and
electrostatic energies are calculated at grid points with 2 Å
spacing, a positively charged carbon atom, and a distance-
dependent dielectric constant with MMFF94 charges.42 The
alignment was also used for CoMSIA steric, electrostatic, and
hydrophobic similarity index fields29 using a probe with a
charge of +1, a radius of +1, a hydrophobicity of +1, and an
attenuation factor R of 0.3 for Gaussian-type distance depen-
dence. Cross-validated analyses48 were run using SAMPLS49
or two and five cross-validation groups with random selection
of group members. PLS (partial least squares)50 analyses using
two or five random cross-validation groups were averaged over
100 runs. For validation, all affinities were randomized51 100
times and subjected to PLS and the mean cross-validated r2
was calculated.
3.3.3. 3D-QSAR Model Validation. Progressive shuf-
fling19,52,53 was used for randomization. Biological activities
were randomized only within 2-12 individual subgroups,
while the relationship between subgroups is not changed. This
allows the direct evaluation of the stability of each model. With
12 subgroups, only a local perturbation of activities is intro-
duced, while for two subgroups, a much larger portion of the
data set is randomized. The number of subgroups was plotted
against the mean cross-validated r2 value per model (20
randomizations each). The 2D fingerprints were generated
using the program UNITY54 for selection of training and test
sets; their similarity is given as the Tanimoto coefficient.55
3.3.4. AFMoC Scoring Function. AFMoC30 implemented
in DrugScore 1.256,57 is used to derive a tailored scoring
function for fXa. Interaction fields for ligands were calculated30
using the fXa/50 complex with 1 Å grid spacing. The grid box
embedded all inhibitors with a margin of at least 4 Å in each
direction. Interaction fields from the following SYBYL atom
type43 were analyzed: C.3, C.2, C.ar, O.3, O.2, N.am, Cl, N.3,
N.ar. A half-width of 0.85 for the Gaussian function to
distribute atomic protein-ligand interactions was used. A
dimensionless value of 10 was used for the height of the
repulsive Gaussian function at the origin of atom-atom
contacts. Analyses with and without leave-one-out cross-
validation were performed using AFMoC’s implementation of
SAMPLS49 and PLS,50 respectively. Because field types from
rarely represented ligand atoms are not included in statistical
analyses, an empirical scaling factor is applied to correct pKi
values by these contributions.30 Consequently, all r2 and r2(cv)
values consider only the part of the binding energy used for
PLS,58 while these correspond to statistical parameters con-
sidering the total binding affinity. Visual inspection of AFMoC
results is based on the contouring of std*coeff fields at
appropriate levels.
4. Results and Discussion
4.1. X-ray Structure Analysis. Crystals of the factor
Xa/inhibitor complexes diffracted reasonably well with
resolutions between 2.1 and 3.3 Å (Table 2). The
inhibitor binding modes are well-defined as much as the
resolution allows. Hence, the 2.1 Å structure is ac-
curately defined, while the 3.3 Å indicates the overall
binding mode. However, high-resolution reference struc-
tures and the high degree of ligand similarity resulted
in more accurate structures than expected without
references.
4.2. Inhibitor Binding Modes: S1 Subsite. Most
fXa inhibitors rely on the interaction of a basic moiety
with Asp189 at the bottom of the protease S1 pocket.
However, six X-ray structures solved at different stages
of this project indicate a favorable nonbasic interaction
with this site. Collectively these X-ray structures reveal
that the inhibitor dichlorophenyl group is invariably
located in S1. Figure 1 summarizes the inhibitor binding
modes of all X-ray structures including structurally
conserved water (cyan spheres) within 4 Å around the
ligand. Here, the fXa binding site is indicated as a
solvent-accessible surface colored by subsite depth.44,47
In particular, the para chlorine atom is involved in a
direct lipophilic contact pointing toward the center of
the aromatic ring of Tyr228 at the back wall of S1
(yellow in Figure 1). Distances from 3.6 to 4.5 Å for this
Cl‚‚‚C.ar (Tyr228) aromatic interaction are observed in
all X-ray structures, in accord with data from knowledge-
based and force field methods. The carbon-chlorine
bond is directed toward the plane of the Tyr228 ring
(dihedral angle C-Cl‚‚‚Tyr228(Centroid)-Tyr228(Cz): 58°).
It appears that the gain in binding energy from this
lipophilic contact could compensate for the lack of a
polar interaction of a basic group to Asp189, resulting
in inhibitors with low nanomolar affinity. This lipophilic
interaction between chlorine and an aromatic ring
(“chloro-binding mode”) is also documented in 41 protein-
ligand X-ray structures in the database ReliBase,61
retrieved using a 3.5-4.5 Å distance between atoms for
3D searching. The interaction of nonbasic groups in the
S1 pocket of the family of Ala190-serine proteases has
previously been described in the search for thrombin32
and fXa inhibitors.23,24 A total of 19 PDB entries with
this interaction were analyzed for factor Xa (PDB
codes: 1ioe, 1iqe, 1iqf, 1iqg, 1iqh, 1iqi, 1iqj, 1iqk, 1iql,
1iqm, 1iqn, 1mq5, 1mq6, 1nfu, 1nfw, 1nfx, 1nfy) and
rat trypsin mutants (1jl7, 1ql9). Even in the presence
of a benzamidine plus a chlorothiophene or chloroben-
zothiophene, the molecules interact using the neutral
group in the “chloro-binding mode”, whereas the S4
pocket accommodates the basic moiety.23 This binding
mode is in line with our X-ray structures, showing that
electrostatic interactions in S1 are not mandatory for
high affinity, while neutral replacements open the route
to inhibitors with increased potential for oral bioavail-
ability.
3.3.5. TScore Scoring Function. In the tailored scoring
function TScore, protein-ligand interactions are evaluated for
residues within 8 Å around the ligand in the fXa/50 complex.31
Descriptors are calculated using an internal C++ implemen-
tation of Chemscore59 and PLP.60 For each residue, the
following descriptors were calculated: Chemscore hydrogen-
bonding and lipophilic terms; PLP steric contact and hydrogen
bonding terms. In addition, the summation over all residues
in a binding site for each of the four terms, corresponding to
original Chemscore and PLP terms, plus the Chemscore
rotatable bond term and total score were included in the
statistical analysis. The 126 relevant descriptors were au-
toscaled and subjected to PLS50 and SAMPLS.49 Mild variable
selection resulted in a significant and predictive TScore
function based on 52 descriptors. TScore visual inspection is
based on coloring residues according to PLS std*coeff values.
This para chlorine atom is situated in an area
occupied by a structurally conserved water molecule in
complexes with benzamidines in S1. This indicates that
solvent displacement is required for binding. The free
energy cost upon displacement of this water molecule
is likely to be close to the maximum value of 2.0 kcal/
mol.62 A view into fXa S1 pockets for a benzamidine-

3298 Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9
Matter et al.
Figure 1. Comparison of X-ray binding modes for different 3-oxybenzamides in complex with human factor Xa. The following
compounds are shown: (A) 57 (153 nM), (B) 50 (18 nM), (C) 37 (51 nM), (D) 63 (650 nM), (E) 87 (1412 nM), (F) 56 (25 nM). The
experimentally determined factor Xa binding site is presented as solvent-accessible surface color-coded by cavity depth (from
orange to blue). Essential pockets are indicated as S1, S4, and EBP (ester-binding pocket). Structurally conserved water is displayed
with cyan spheres.
containing ligand (Figure 2A, PDB code 1lqd19) in
comparison to 50 (Figure 2B) with a chlorine-Tyr228
interaction illustrates this different binding interaction
in S1.
the ester binding pocket (EBP) on top of the Cys191-
Cys220 disulfide bridge and surrounded by Gln192,
Arg143, and Glu147 side chains adjacent to S1.21
Figure 1A shows the structure of 57 in fXa with an
arginine moiety in S4 as the start for lead optimization
(Ki(fXa) ) 153 nM, resolution of 2.1 Å). The 4-methoxy
group located in the EBP is in contact with Glu147 and
Gln192 Cγ atoms and the main chain carbonyl of
Glu147. The C-terminal arginine diethylamide does not
favorably interact and is disordered in the electron
density maps. The arginine side chain is accommodated
in S4 lined by the aromatic side chains of Tyr99,
Phe174, and Trp215. The guanidine group is located in
the “cation hole” at a position that is occupied by a
pyridyl nitrogen in most structures. The terminal
guanidine nitrogens interact with a water molecule; one
of them is hydrogen-bonded to the inward-pointing
backbone carbonyl oxygens of Glu97 and Thr98. The
side chain of Glu97 is partly disordered and does not
interact directly with the guanidine, while neighboring
negative charges of Glu97 and Asp100 create a region
favorable for the binding of positively charged residues.
This distal part of S4 seems to flexibly adapt to the
requirements of the ligand. The ligand arginine carbonyl
oxygen interacts via a hydrogen-bonding network to two
structurally conserved water molecules with the polar
side chains of Arg222 and Glu217.
This experimental preference for a favorable interac-
tion involving Tyr228 is also uncovered in molecular
interaction fields computed using GRID and an organic
chlorine as probe atom.39 In addition, favorable contour
regions were also obtained using the complementary
knowledge-based SuperStar approach.63 Both interac-
tion regions are highlighted by green contours in parts
C and D of Figure 2 from profiling the binding site of
the fXa/50 complex using GRID and SuperStar, respec-
tively. The para chlorine atom in our X-ray structures
is also closer to more polar atoms at the edge of this
pocket, namely, Trp215-CdO (3.6 Å) and Ile227-NH
(3.4 Å), while the Asp189 carboxylate is at a distance
of 4.5 Å.
The ortho chloro atom is more solvent-exposed and
interacts with the Oγ group of the inward-directed
Ser195 and a conserved water. Asp189 at the bottom of
S1 points toward the aromatic ring and is not involved
in a favorable interaction.
4.3. Inhibitor Binding Modes: Central Scaffold
and S4 Subsite. In general the position of the 3-oxy-
benzamide aromatic ring is well-defined in electron
density maps with the benzamide nitrogen hydrogen-
bonded to the carbonyl oxygen of Gly216. The small
oxybenzamide 4- and 5-substituents are oriented toward
The X-ray structure of 50 (Ki(fXa) ) 18 nM, resolution
of 2.7 Å) reveals a similar binding mode with the

Factor Xa Inhibition
Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9 3299
Figure 2. Comparison of factor Xa S1 binding modes for a typical benzamidine containing ligand (PDB code 1lqd) (A) and compound
50 (B) with a chlorine-Tyr228 interaction to illustrate different interaction pattern. Selected S1 amino acids are shown with
solvent-accessible binding site surface colored by cavity depth. Depth cueing is used for a view into S1. Shown is the detailed
analysis of the fXa/50 complex for favorable interactions with organic chlorine probes using GRID (C) and SuperStar (D). Favorable
interaction regions are consistently highlighted by green contours.
4-methoxy group situated in the ester binding pocket
and the favorable Cl-Tyr228 interaction in S1. This
structure is shown in Figures 1B and 3 with key
interactions highlighted. The electron density of the
dichlorophenyl and piperidylpyridyl substituents in S1
and S4 is well-defined, while the weak density in the
central part indicates some disorder at the scaffold. The
benzamide nitrogen interacts with Gly216-CO, while
the piperidine in its chair conformation orients the
pyridine substituent perfectly into S4, stacking it be-
tween the aromatic side chains of Tyr99 and Phe174.
The pyridine nitrogen, which corresponds closely to one
guanidine nitrogen in the complex of fXa/57, is hydrogen-
bonded to the carbonyl oxygen of Thr98 via a structur-
ally conserved water molecule in S4, which also inter-
acts with the carbonyl oxygen of Ile175 and the Thr98
hydroxyl group (Figure 3).
The binding mode of 37 in fXa, shown in Figure 1C
(Ki(fXa) ) 51 nM, resolution of 2.5 Å), is identical, within
experimental error, to 50 with clear and unambiguous
electron density. The 5-amide substitution is directed
toward the ester-binding pocket, while none of the four
polar neighboring side chains are closer than 3.8 Å to
this group. The 5-amide group shows one direct hydro-
gen bond with a structurally conserved water molecule,
while the amide oxygen and nitrogen atoms are indis-
tinguishable in the electron density maps. Conse-
quently, for 3D-QSAR studies, all derivatives were
consistently built from the actually shown orientation.
Figure 1D shows the binding mode of the inhibitor
63 (Ki(fXa) ) 659 nM, resolution of 3.1 Å). Despite its
unambiguous electron density, the resolution of 3.1 Å
does not allow us to determine atomic positions with
extreme accuracy. However, the binding mode of this

3300 Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9
Matter et al.
tuted piperazine located above Gly216 without the
possibility of any favorable hydrogen bond interaction
with Gly216-CO. Again, the electron density of the
bound inhibitor is clear but indicates some conforma-
tional heterogeneity. The main interactions in S4 are
not possible with the N-dimethylacetamide substituent
stacked between Phe174 and Tyr99, situated at the
position of the pyridine ring in 50. Consequently the
replacement by a 4-piperidylmethyl substituent results
in a slight increase of binding affinity (62, Ki(fXa) ) 650
nM).
Finally Figure 1F displays the binding mode of 56 (Ki-
(fXa) ) 25 nM, resolution of 3.3 Å), deduced from a
weakly diffracting crystal, thus limiting the accuracy
of the structure. However, the electron density of the
inhibitor is unambiguous and indicates a similar bind-
ing mode to 50. This compound lacks the ortho chlorine
atom in S1 and has two hydrophobic substitutions in
the EBP. From the data set, it can be concluded that
the chlorophenyl group is inserted ∼0.7 Å less deep into
the S1 pocket, while this difference is within the error
margins. However, this is consistent with docking,
suggesting an influence of the size of the EBP substitu-
tion to binding into S1, probably because of contacts to
the bulky Cys191-Cys220 disulfide bond. This com-
parative analysis supports the assumption of a common
binding mode to explain the structure-activity relation-
ship.
4.4. CoMFA Model and Validation Studies. All
3D-QSAR models based on the fXa-derived alignment
rule and different force field minimizations (QXP,
MMFF94s) showed a high degree of consistency. During
this project, preliminary 3D-QSAR models with fewer
compounds at a particular time were used in each
design cycle and then updated, while here, we present
a final model encompassing a larger degree of structural
variations in the set of 3-oxybenzamides (cf. SAR set 1
in Table 1). By use of a 2 Å grid spacing, a CoMFA model
with an r2(cv) value of 0.714 for six PLS components
and a conventional r2 of 0.947 were obtained (model A,
Table 3 and Figure 4).
The graph of observed versus fitted biological activi-
ties for model A is displayed in Figure 4. The steric field
descriptors explain 47% of the variance, while the
electrostatic field accounts for 53%, from the normalized
sum of standard deviations after CoMFA_STD scaling
multiplied by PLS coefficients from the final non-cross-
validated PLS model. From MMFF94s minimizations,
a CoMFA model with a lower r2(cv) value of 0.585 for
six PLS components and a conventional r2 of 0.913
resulted (model C, Table 3). Model A was extensively
validated, which underscored its predictive power and
significance. All affinity predictions obtained from this
and subsequent models are listed in Table 4.
The effect of the alignment relative to the grid
definition was evaluated by consistently moving all
compounds in increments of 0.5 Å in all three dimen-
sions x, y, and z without affecting the superposition. The
r2(cv) values range from 0.636 to 0.758 (mean r2(cv) of
0.70; SD ) 0.04), suggesting only a minor dependence
on the orientation of the grid. A variation of 14 atom
types as probe atom for CoMFA produced r2(cv) values
from 0.714 to 0.744 with Cl or carbon-based atom types
showing r2(cv) values of ∼0.72 and nitrogen and oxygen
Figure 3. Details of the fXa/50 complex (Ki ) 18 nM) from
the 2.7 Å X-ray structure: (A) detailed interactions between
50 and the fXa binding site with key residues indicated; (B)
same view with solvent-accessible surface colored by cavity
depth; (C) two-dimensional representation of essential protein-
ligand interactions.
inhibitor with the pyridyl moiety replaced by a lipophilic
isopropyl substituent directed toward the edge of S4 is
similar to that of the fXa/50 complex without any
unfavorable interaction, indicating that the general
orientation is not influenced by S4 modifications.
Another binding mode for a weak inhibitor is given
in Figure 1E for 87 (Ki(fXa) ) 1410 nM, resolution of
2.65 Å). Here, the S4 directed substituent is a substi-

Factor Xa Inhibition
Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9 3301
Table 3. Model Statistics for 3D-QSAR and Tailored Scoring Functions
alignment
model
no. compds
r2(cv)
SD
PLS components
r2
pred r2
(a) 3D-QSAR Models To Explain Factor Xa Activity of 3-Oxybenzamidesa,b
CoMFAc
CoMFA
CoMFAc
CoMFA
CoMSIAc
CoMSIA
CoMSIAc
CoMSIA
QXP
A
B
C
D
E
F
80
107
80
107
80
107
80
0.741
0.730
0.585
0.616
0.609
0.664
0.600
0.612
0.507
0.484
0.610
0.580
0.587
0.542
0.607
0.583
6
6
6
6
5
6
8
6
0.947
0.889
0.913
0.874
0.898
0.899
0.940
0.882
0.732
0.696
0.784
0.679
QXP
MMFF
MMFF
QXP
QXP
MMFF
MMFF
G
H
107
(b) Tailored Scoring Functions To Explain Factor Xa Activity of 3-Oxybenzamidesd
AFMoCe
AFMoCe
TScoref
TScoref,g
TScoref
QXP
I
80
80
80
80
80
0.495
0.414
0.572
0.644
0.343
0.721
0.716
0.615
0.561
0.767
4
4
5
5
6
0.723
0.659
0.790
0.827
0.775
0.577
0.741
0.301
0.399
0.222
MMFF
QXP
J
K
L
M
QXP
MMFF
a CoMFA and CoMSIA models were derived using a minimum σ of 2. r2(cv): cross-validated r2 using leave-one-out. SD: standard
deviation of error from leave-one-out PLS model. PLS components: optimal number of components. r2: non-cross-validated regression
coefficient. b Models based on 80 compounds were derived using SAR set 1 in Table 1. c Models subjected to extensive statistical validation.
d r2(cv): cross-validated r2 using leave-one-out. SD: standard deviation of error from leave-one-out PLS model. PLS components: optimal
number of components. r2: non-cross-validated regression coefficient. e AFMoC models were derived using nine different field types: C.3,
C.2, C.ar, O.3, O.2, N.am, Cl, N.3, N.ar and 80 compounds from SAR set 1 in Table 1. f TScore models were derived using Chemscore
lipophilic, hydrogen bond terms, PLP hydrogen bond, contact terms for each residue within 8 Å around the inhibitor binding site plus
global Chemscore and PLP terms for 80 compounds from SAR set 1 in Table 1. g TScore model selecting 51 informative variables based
on PLS coefficients from model K.
analysis without cross-validation led to a model for
predicting the remaining compounds to produce the
predictive r2 value for each model. The r2(cv) value is
low with less than 44 diverse compounds in the training
set. Although the conventional r2 is high, the predictive
ability is not sufficient. The mean Tanimoto coefficient
for this 44-compound training set is 0.83, while the pair
of most similar compounds in this subset shows a
Tanimoto coefficient of 0.88. When the subset size is
increased, the cross-validated r2 reaches values between
0.55 and 0.78. For the corresponding external predic-
tions the predictive r2 is between 0.67 and 0.85, dem-
onstrating stable models with good predictive capabili-
ties for these external test sets. This example suggests
that a reliable prediction of biological affinities can be
Figure 4. Graph of observed versus fitted binding affinities
expected for novel molecules with a higher similarity
than defined by this Tanimoto threshold of ∼0.83. The
degree of extrapolation increases with decreasing simi-
larity, thus leading to less reliable affinity predictions
for external test sets.
for the final non-cross-validated CoMFA model A.
probes r2(cv) of ∼0.74, suggesting only a slight depen-
dence on the probe (mean r2(cv) of 0.73; SD ) 0.01). One-
hundred randomizations of biological activity resulted
in a mean r2(cv) of -0.05 (SD ) 0.06; high, 0.10; low,
-0.21), showing that model A is significantly better than
a random model.
Although cross-validation reflects the predictive power
of a model, the leave-one-out method might produce too
optimistic r2(cv) values. Thus, PLS analyses were run
100 times with two and five randomly chosen cross-
validation groups containing randomly selected com-
pounds for prediction. The mean r2(cv) value of 0.683
for six PLS components and five cross-validation groups
(SD ) 0.03; high, 0.742; low, 0.604) is slighly lower than
using the leave-one-out method. By use of only two
cross-validation groups, a lower mean r2(cv) value of
0.610 was observed (SD ) 0.06; high, 0.722; low, 0.450).
Progressive shuffling19,52,53 was applied to assess
model stability against variations in biological data.
After the data set is divided into 2-12 groups using
activity thresholds, each group is internally randomized.
By use of 12 groups, only a small uncertainty is probed
resulting in a mean r2(cv) of 0.68, while for two
subgroups, a larger portion of the y-block is randomized.
This interpretation aids in understanding the effect of
biological errors on PLS predictions and estimates
model stability. With one group, the results are similar
to complete randomization; negative r2(cv) values are
observed. Shuffling with more than two subgroups
produces models of remarkable quality (mean r2(cv) for
three groups 0.53), suggesting that the final models are
less dependent on minor biological variations. These
extensive validation studies collectively support the
finding of a stable, significant, and predictive PLS
model.
The influence of increasing information in X-space on
model predictivity is estimated from incrementally
dividing the set of 80 3-oxybenzamides into training and
test sets using statistical design,64 producing subsets
with 20-76 members. For each subset a leave-one-out
PLS analysis served to extract the r2(cv), while a PLS
4.5. CoMSIA Model and Validation Studies. Simi-
lar statistical results were obtained using CoMSIA
steric, electrostatic, and hydrophobic fields. By use of a

3302 Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9
Matter et al.
Table 4. Fitted and Predicted Factor Xa Affinities from 3D-QSAR Models for 3-Oxybenzamidesa
ID SAR_set FXa_pKi CoMFA_A CoMSIA_E AFMoC_I TScore_L ID SAR_set FXa_pKi CoMFA_A CoMSIA_E AFMoC_I TScore_L
1
2
1
1
2
1
1
1
2
1
1
1
2
2
1
1
1
1
2
2
1
1
2
2
1
1
1
1
1
2
1
1
1
2
2
1
1
1
1
1
2
2
2
2
2
1
2
2
1
1
1
1
1
1
2
2
4.70
4.30
4.40
4.43
4.44
3.99
3.91
3.50
2.81
2.71
2.60
3.38
2.34
2.00
4.15
3.69
3.69
3.51
2.34
3.15
2.48
2.81
2.00
2.00
3.97
4.12
3.81
4.43
3.85
4.36
4.54
4.32
4.12
3.50
2.75
3.23
4.29
4.55
4.21
3.63
3.90
4.30
3.58
4.41
4.70
4.07
2.98
5.00
4.89
4.74
4.39
4.27
4.39
4.89
4.66
4.05
4.23
4.47
4.26
4.09
4.08
3.40
2.58
2.69
2.55
3.79
2.52
1.98
4.23
4.00
4.13
3.35
2.53
2.70
2.59
3.68
1.99
1.96
3.98
4.25
4.04
3.74
4.04
4.33
4.08
3.33
3.22
3.14
2.78
3.28
4.42
4.35
3.69
3.65
4.57
3.50
3.86
3.84
4.89
4.84
3.03
5.00
4.51
4.58
4.58
4.40
4.05
4.26
4.78
4.08
4.38
4.71
4.73
4.04
4.03
3.22
2.63
3.12
2.41
3.54
2.51
2.07
4.57
3.89
4.16
3.10
2.51
3.02
2.27
3.40
2.26
2.06
3.87
4.18
4.22
3.80
3.63
4.01
3.64
3.80
3.28
3.41
3.31
3.27
4.21
4.21
3.90
3.71
4.40
3.63
4.22
3.93
4.73
4.83
3.08
4.72
4.35
4.20
4.37
4.51
4.08
4.55
5.18
4.04
4.32
4.53
3.29
5.23
4.67
3.63
3.20
2.92
2.34
4.39
2.99
1.66
2.78
4.88
4.72
3.57
3.05
2.81
2.21
4.16
2.42
1.90
3.96
3.32
2.99
3.62
4.33
4.15
3.97
3.34
3.95
3.05
3.15
3.78
4.07
4.45
3.66
3.61
4.10
3.11
3.36
3.42
3.63
3.66
3.06
4.94
3.96
3.67
3.87
3.99
4.49
4.84
4.69
3.83
4.01
4.31
4.50
4.30
4.07
3.63
2.73
2.84
2.91
2.60
2.41
2.15
4.25
4.12
3.40
3.74
2.64
2.77
2.94
2.44
1.65
1.96
3.88
4.36
4.26
3.84
4.11
3.88
3.74
3.03
3.37
3.63
2.76
3.11
3.96
4.02
2.42
2.64
3.94
2.48
2.64
3.25
4.42
4.14
2.84
4.38
3.97
4.24
4.31
4.24
4.37
4.26
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
1
2
1
1
1
1
1
2
2
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
2
1
2
1
1
1
1
2
1
1
1
1
1
1
1
1
1
2
1
1
1
1
3.55
4.60
3.82
3.72
4.24
3.96
3.58
3.22
3.19
3.02
3.01
2.95
2.81
3.02
3.39
2.21
2.00
2.00
3.76
2.29
3.85
2.00
2.00
2.00
2.00
2.83
2.86
2.90
2.00
2.00
2.00
2.36
2.85
2.00
2.98
2.39
2.24
3.78
3.51
2.00
2.27
2.52
2.66
2.41
2.65
2.37
2.03
2.19
2.21
2.60
2.59
2.55
2.31
3.34
4.89
3.83
3.59
4.04
4.09
3.68
3.35
3.20
3.32
3.06
3.32
2.72
3.18
3.13
2.45
2.32
2.44
3.93
2.38
3.70
2.21
1.66
1.89
1.73
3.15
3.21
3.09
2.00
2.06
2.09
2.42
2.81
2.32
2.90
2.37
2.43
3.64
3.41
1.96
2.27
2.28
2.75
2.33
2.82
2.50
1.72
2.02
1.86
2.68
2.60
2.34
2.62
3.88
4.97
4.21
3.93
3.68
4.02
3.34
3.27
2.61
3.44
2.83
3.09
2.95
2.84
2.85
2.69
2.28
2.51
3.84
2.25
3.71
1.87
1.99
1.92
1.81
3.11
2.78
3.16
2.24
1.85
2.15
1.97
3.08
2.05
2.94
2.13
2.69
3.49
3.40
1.89
2.45
1.89
2.72
2.43
2.56
2.46
2.28
1.98
1.68
2.44
2.63
2.33
2.46
3.27
4.97
3.39
3.30
3.75
3.40
3.52
3.11
2.92
3.56
2.64
2.61
2.80
2.32
2.68
2.67
2.63
3.05
3.64
2.95
3.52
1.70
2.14
1.61
2.30
2.88
3.13
3.09
2.31
2.62
2.08
2.75
2.82
2.39
2.61
2.43
2.65
3.04
3.15
1.88
2.34
2.06
2.79
2.09
2.87
2.31
2.16
2.37
2.42
2.69
2.92
2.61
4.11
4.23
4.43
3.67
3.52
4.09
4.18
3.95
3.21
3.68
3.70
2.85
2.98
3.20
2.86
3.32
2.18
1.86
2.43
3.91
2.16
3.79
1.86
2.36
2.13
2.06
2.82
3.16
2.92
2.47
2.56
2.22
2.81
3.06
2.47
2.87
2.03
2.46
2.93
2.69
2.22
2.36
2.42
2.80
2.01
2.68
2.28
2.09
2.11
1.79
2.89
2.59
2.61
3.75
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
a Experimental biological activity pKi_fXa is expressed as log[(1/K )(1000)]. See text and Table 3 for details on 3D-QSAR models and
i
tailored scoring functions. Predictions for all compounds from models A, E, I, L, and SAR set 1 were based on the corresponding non-
cross-validated PLS models with optimal number of components, while compounds indicated as SAR set 2 were used as the external
prediction set.
2 Å grid spacing, a CoMSIA model with an r2(cv) value
of 0.609 for six PLS components and a conventional r2
of 0.898 was obtained (model E, Table 3). The steric field
descriptors explain 16% of the variance, the electrostatic
descriptors explain 48%, and the additional hydrophobic
field explains the remaining 36%. The alignment from
MMFF94s minimization produced a CoMSIA model
with a comparable r2(cv) value of 0.600 for eight PLS
components, and a conventional r2 of 0.940 resulted
(model G, Table 3). Model E was subjected to validation,
which collectively supports the finding of a stable and
predictive model.
mean r2(cv) of -0.10 (SD ) 0.12; high, 0.21; low, -0.47)
is observed. When over 100 PLS analyses with two
random cross-validation groups are averaged, a mean
r2(cv) of 0.464 (SD ) 0.08; high, 0.639; low, 0.228)
results. This value is increased for five cross-validation
groups. Here, a mean r2(cv) value of 0.553 for six PLS
components results (SD ) 0.04; high, 0.640; low, 0.448).
The incremental splitting of the data set in training and
test using statistical design led to comparable results
to CoMFA. For CoMSIA, the first reliable PLS models
emerge with 48 training set molecules, a r2(cv) of.0.57,
and a predictive r2 of 0.63. In general, both the cross-
validated and predictive r2 values are slightly lower for
CoMSIA. This study based on external test sets suggests
that for this case the CoMFA models have a higher
No effect of the alignment relative to the grid was
observed because of the CoMSIA Gaussian-type smooth-
ing function.65 When the activities are randomized, a

Factor Xa Inhibition
Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9 3303
Figure 5. Experimental versus fitted or predicted binding affinities for 3D-QSAR models (Table 3). Shown are 80 training
compounds as dots, while crosses represent 27 compounds for external predictions: (A) CoMFA model A; (B) CoMSIA model E.
predictive power than CoMSIA, while CoMSIA is less
dependent on alignment inaccuracies and might be
comparable from a practical point of view for design.
For each training set with acceptable statistical param-
eters it was checked whether its interpretation in terms
of CoMFA and CoMSIA contour maps led to similar
conclusions, which was always the case. Progressive
shuffling applied to the final CoMSIA model also
demonstrates its robustness against variations in bio-
logical data. The increase of the mean r2(cv) values with
more than one subgroup is obvious, suggesting that
shuffling using more than three subgroups still produces
models of remarkable quality for CoMSIA (mean r2(cv)
of 0.59 for 3 subgroups and 0.69 for 12 subgroups). This
analysis again shows that both QSAR techniques are
robust and tolerate minor biological variations.
4.6. Prospective Design Based on 3D-QSAR Mod-
els. The final 3D-QSAR models were applied prospec-
tively to 27 3-oxybenzamides as an external set (SAR
set 2 in Table 1). New compounds were docked and
aligned in a consistent way and minimized using the
appropriate force field. The obtained predictive r2 values
for models in Table 3 are 0.732 and 0.696 for CoMFA-
QXP (model A) and CoMFA-MMFF (model C), respec-
tively, and 0.784 and 0.679 for CoMSIA-QXP (model E)
and CoMSIA-MMFF (model G), respectively. A graph
of experimental versus predicted fXa binding affinities
is given in parts A and B of Figure 5 for models A and
E, while affinity predictions for training and test sets
are summarized in Table 4 for all relevant 3D-QSAR
models and tailored scoring functions discussed below.
The inspection of Figure 5 suggests that the final 3D-
QSAR models are of comparable statistical quality in
terms of predicting novel binding affinities with a
slightly better performing alignment derived from dock-
ing using QXP. However, this minor difference might
also be due to the partial side chain flexibility intro-
duced in QXP docking. Both CoMFA and CoMSIA
models were applied to prioritize synthesis candidates.
Prospective predictions for 44 additional new com-
pounds resulted in a predictive r2 value of 0.68 for
CoMFA model A and 0.82 for CoMSIA model E (no data
given), clearly showing that the quality of external
predictions was acceptable for both CoMFA and CoM-
SIA. Indeed, this model was used for further design
cycles. A combination of structure-based design with
inspection of 3D-QSAR contours proved very effective
in supporting the design of novel inhibitors. Synthesis
proposals were generated, evaluated using this proce-
dure, and reliably ranked by 3D-QSAR affinity predic-
tions. Each design cycle then was completed by genera-
tion and applications of novel, improved 3D-QSAR
models. This stepwise design and synthesis procedure
based on CoMFA predictions constantly improved sta-
tistical results and predictivity in this series.
The final 3D-QSAR models encompassed all 107
3-oxybenzamides from Table 1 within a single analysis,
which did not significantly alter the statistical quality.
The corresponding CoMFA and CoMSIA models based
on QXP or MMFF alignments are summarized in Table
3 with r2(cv) values of 0.730 and 0.616 for CoMFA and
0.664 and 0.612 for CoMSIA, respectively. No significant
changes in CoMFA and CoMSIA contour regions were
obvious compared to models A and E.
4.7. Tailored Scoring Functions. The 3D-QSAR
models were very consistent with each other and with
the two tailored scoing functions, despite the use of
different force field protocols. Note that the QXP align-
ments were statistically superior in all cases, however.
By use of a 1 Å grid spacing, an AFMoC model with an
r2(cv) value of 0.495 for four PLS components and a
conventional r2 of 0.723 was obtained (model I, Table 3
and Figure 9). Its graph of observed versus fitted
biological activities is displayed in Figure 6A. Dots
indicate training set molecules (SAR set 1 in Table 1),
and crosses represent the test set (SAR set 2 in Table
1). In total, nine atom-type-derived interaction fields
were utilized to derive this model: C.3, C.2, C.ar, O.3,
O.2, N.am, Cl, N.3, N.ar. Because the atom types occur
frequently in test and training sets, the number of fields
in the final model was not reduced. Changing the
minimization protocol led to a PLS model with a slightly
lower r2(cv) value of 0.414 for four PLS components and
a conventional r2 of 0.659, while the predictive r2 for
the external test set (SAR set 2) was improved from
0.577 (AFMoC model I, QXP) to 0.741 (AFMoC model
J, MMFF) (see Table 4). The relative contributions of
individual field types from both AFMoC models are
presented in Table 5, showing that lipophilic atom types
such as C.3, C.ar, and Cl dominate the models. Prospec-

3304 Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9
Matter et al.
Figure 6. Experimental versus fitted or predicted binding affinities for different scoring functions: (A) AFMoC model I; (B)
TScore model L; (C) DrugScore global scoring function; (D) ChemScore global scoring function.
Table 5. Relative Contributions of Field Types to the Final
residue plus global terms), while a mild variable selec-
tion reduced this number to 51 relevant terms listed in
Table 6. The final TScore model L then showed an r2(cv)
value of 0.644 for five PLS components and a conven-
tional r2 of 0.827 (model L; Table 3b and Figure 10).
The graph of observed versus fitted biological activities
for this model is displayed in Figure 6B. Changing the
minimization protocol resulted in a PLS model with a
lower r2(cv) value of 0.343 for six PLS components, a
conventional r2 of 0.767, without significant prediction
for the external test set (predictive r2 of 0.222 for SAR
set 2). In contrast, the previous model L showed a
predictive r2 value of 0.399 for this external set (Table
4) and a predictive r2 value of 0.72 in a prospective study
of 44 additional compounds (no data given). While the
predictive r2 value for SAR set 2 of model L is lower
than observed for the 3D-QSAR models (cf. 0.732 for
CoMFA model A), this value is still acceptable given the
chemical diversity of this test set. A detailed inspection
of predictions for both AFMoC and TScore highlights
intrinsic features of scoring functions. Those tend to
extract SAR information close to the protein surface,
while 3D-QSAR are able to capture substitution effects,
which are more distant to the surface but relevant to
affinity for other reasons, such as entropy, long-range
electrostatic effects, and others. Because this informa-
tion is explicitly excluded from both tailored scoring
functions, 3D-QSAR approaches still offer complemen-
AFMoC Modelsa
% contribution
field type
AFMoC_I
AFMoC_J
C.3
C.2
22.4
1.7
22.6
2.6
C.ar
O.3
O.2
N.am
Cl
N.3
N.ar
41.3
4.6
38.2
5.4
0.7
1.2
1.1
1.8
20.7
0.3
16.9
0.6
7.2
10.5
a PLS-derived relative contributions of different AFMoC field
types for the nine different fields and AFMoC models I and J.
tive predictions for 44 additional compounds resulted
in predictive r2 values of 0.69 for model I and 0.75 for
model J (no data given), showing again that the quality
of external predictions for these tailored scoring func-
tions is comparable to that of CoMFA and CoMSIA.
TScore as an alternative tailored scoring function also
led to predictive PLS models, summarized in Table 3b.
By use of descriptors for binding site residues within 8
Å around the template molecule 50, a TScore model with
an r2(cv) value of 0.572 for five PLS components and a
conventional r2 of 0.790 were obtained after QXP
minimization (model K, Table 3b). This model encom-
passed all 126 nonzero TScore descriptors (four per

Factor Xa Inhibition
Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9 3305
Table 6. PLS Coefficients of Factor Xa Amino Acid Properties
with the inhibitor 48 (Ki(fXa) ) 10 nM) having a
4-piperidylpyridyl moiety in S4 and a dichlorophenyl
substituent in S1. Green contours in Figure 7A (>80%
contribution) indicate regions where steric bulk is
favorable for fXa affinity, while yellow contours (<20%
contribution) highlight regions where bulky substituents
are detrimental. A similar analysis is provided for the
electrostatic SD*coeff field (Figure 7B). Blue contours
(>80% contribution) refer to regions where an increase
of positive charge (or less negative charge) is favored to
enhance affinity, while red contours (<20% contribution)
indicate regions where an increase of negative charge
is favorable for biological affinity. Although 3D-QSAR
results are derived from ligand information, these
contour maps are consistent with steric, electrostatic,
and hydrophobic requirements of the fXa binding site,
as indicated in parts C and D of Figure 7.
Steric bulk at positions 4 and 5 of the benzamide ring
is favorable for affinity. Here, a chlorine atom is
optimally situated within the ester-binding pocket, as
indicated by the green contour on top of the Cys191-
Cys220 disulfide bridge, which is limited in size by the
near Gln192 side chain. The neighboring yellow contour
indicates the optimal size of this pocket and suggests
clashes with the backbone carbonyl of Glu147 and the
Arg143 side chain if the size of substituents in the 4- or
5-position is increased. The green contour next to the
5-methoxy substituent of 48 indicates favorable steric
contributions. Although this region opens to the bulk
solvent, some interactions with the side chains of
Glu143 and Arg222 might be possible. The alkoxy group
might also serve to shield the hydrophobic edge of the
aromatic ring from salvation. The main favorable steric
effect in S1 is indicated by a green contour next to the
ortho-chlorine substituent, pointing toward the active
triad Ser195-Oγ and Gln192-CR. The second green
contour in S1 points toward the polar Arg189 and
indicates that some volume of the S1 pocket has to be
occupied to achieve nanomolar affinity even with neu-
tral inhibitors. The optimal fit of the S4 directed
4-piperidylpyridyl substituents is indicated by the green
contour close to the distal ring. The aromatic portion of
this substituent is involved in a favorable π-stacking
contact with the aromatic residues of Phe174 and
Trp215 lining this pocket. The yellow contour in the
vicinity of the terminal pyridyl ring indicates the edge
of S4. This contour is located at the position of one
structurally conserved water molecule within this pocket
and suggests that replacement of this hydrogen-bonded
water molecule in S4 might not be very favorable for
affinity. Hence, extending the size of S4 directed sub-
stituents is detrimental for affinity. Interestingly, the
protein topology of the aryl-binding pocket for this
ligand series consistently differs from protein structures
experimentally observed for other scaffolds23 and is
more similar to 1HCG,22 where fXa is complexed with
a second fXa molecule. In other X-ray structures, this
S4 site is smaller because of a movement of Tyr99
toward Phe174 at its other edge.23
from TScore Model La
Global_terms
coeff
Rot
-0.085
0.084
sum_HB
sum_LI
0.000
sum_PLP_ST
sum_PLP_HB
ChemScore
0.000
0.069
-0.007
residue
HB
LI
PLP_ST
PLP_HB
THR95
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.213
0.000
0.000
0.000
0.000
0.000
0.000
0.000
-1.992
-0.085
-0.024
0.005
0.000
5.381
0.037
0.000
0.622
0.000
-0.102
0.000
0.000
0.000
0.000
0.247
0.000
-0.127
2.884
0.000
0.000
0.000
-0.049
0.073
0.087
0.000
-0.221
0.042
0.000
0.000
-1.717
0.000
0.000
0.000
0.219
0.000
0.000
0.000
0.080
0.000
0.000
0.000
0.000
0.000
0.463
0.000
-0.262
0.000
0.000
0.000
0.000
0.000
-1.280
0.214
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
GLU97
THR98
0.000
TYR99
-0.007
0.014
0.709
-0.022
0.089
ARG143
HIS145
GLU147
LYS148
SER173
PHE174
ILE175
GLU188
ASP189
CYS191
GLN192
GLY193
ASP194
SER195
ILE212
VAL213
SER214
TRP215
GLY216
GLY219
CYS220
ALA221
ARG222
ILE227
TYR228
THR229
-0.124
0.004
0.007
0.283
0.000
0.035
0.011
0.000
-0.154
0.073
0.566
-0.009
-0.055
0.002
-0.009
0.007
0.040
-0.248
-0.018
0.000
-0.002
-9.236
a PLS coefficients from five component model using 52 variables.
HB: Chemscore hydrogen bond term. LI: Chemscore lipophilic
term. PLP_ST: PLP steric contact term. PLP_HB: PLP hydrogen
bonding term. Rot: global Chemscore rotatable bond term. sum-
_HB: global Chemscore hydrogen bond term. sum_LI: global
Chemscore lipophilic term. sum_PLP_ST: global PLP steric
contact term. sum_PLP_HB: global PLP hydrogen bonding term.
Chemscore: global Chemscore total score.
tary insights into prerequisites for binding affinity. Such
a difference between 3D-QSAR and tailored scoring
functions should be more pronounced for inhibitors
binding in shallow, solvent-exposed sites, while signifi-
cant differences might not be apparent for more buried
ligand series. This agrees with internal observations for
other proteins.
For both tailored scoring functions AFMoC and
TScore, the comparison with affinity predictions from
their parent functions DrugScore and Chemscore, shown
in parts C and D of Figure 6, highlights the significant
improvement achieved by adapting those global scoring
functions to the specific environment of a particular
protein binding site. Of course, this strategy requires a
balanced set of active inhibitors, which is not often
available at the start of a new project. Global scoring
functions are valuable in early phases, while tailoring
of scoring terms might offer advantages for lead opti-
mization.
The interpretation of the CoMSIA model is similar
to CoMFA and is also in agreement with the factor Xa
binding site topology. The contour maps for steric fields
are displayed in Figure 8 without (parts A-C) and with
(parts D-F) the solvent-accessible binding site surface.
4.8. Comparing 3D-QSAR Results with the fXa
Binding Site. The steric and electrostatic std*coeff
fields for the CoMFA analysis A are shown in Figure 7

3306 Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9
Matter et al.
Figure 7. Contour maps from CoMFA model A in combination with compound 48 (10 nM). Hydrogens are omitted for clarity. (A)
Shown is the steric std*coeff contour map. Green contours (>80% contribution) refer to regions where additional steric bulk is
favorable for affinity, while yellow contours (<20% contribution) indicate disfavored areas. (B) Shown is the electrostatic std*coeff
contour map. Blue contours (>80% contribution) refer to regions where negatively charged substituents are unfavorable for affinity.
Red contours (<20% contribution) indicate regions where negatively charged substituents are favorable. (C) Same as (A) with
factor Xa solvent-accessible surface to show complementarity to protease topology. (D) Same as (B) with factor Xa solvent accessible
surface.
Again, green contours indicate regions where steric bulk
is favorable, while yellow contours highlight regions
where steric bulk is detrimental for affinity. In addition
to CoMFA, the positive influence of the oxybenzamide
scaffold on affinity is highlighted by an extended green
region surrounding the central part of the fXa inhibitors.
This different interpretation results from the different
information captured in both field types. In CoMFA
force-field-derived molecular interaction fields, interac-
tions of the ligand with potential partners are indicated
outside the volume of this ligand. In contrast, CoMSIA
works by identifying similarities between individual
ligand atoms and substructures within the volume of
the ligands. Hence, the CoMSIA interpretation is fo-
cused on the actual ligand position, while CoMFA
projects these favorable interactions toward the protein
binding site. Consequently the size of the yellow con-
tours in CoMSIA is a better indicator of the size of
particular substituents within S4 and the ester binding
pocket without significantly reducing binding affinity.
The interpretation of the electrostatic contour regions
for both CoMFA and CoMSIA (Figures 7B,D and 8B,E)
shows that an increase in negative charge is preferable
at position 4 of the benzamide scaffold within the ester-
binding pocket on top of the Cys191-Cys220 disulfide
bridge, which agrees with the observation that halogen
substituents in this area increase affinity binding. The
neighboring blue contour indicates a favorable influence
of more positive charge at the edge of this pocket, due
to the interaction of appropriate polar substituents with
the backbone carbonyl oxygen of Glu147. While no
significant influence of electrostatic interactions within
S1 are observed, these forces significantly contribute to
the structure-activity relationship within the S4 pocket.
Here, more positive charge is favorable in the blue
region surrounding the distal pyridyl ring, which is in
agreement with the possibility of a cation-π interaction.
Another blue contour region is observed close to the
piperidyl nitrogen atom, directly at the optimal position
for a favorable interaction of a cation to the Trp215
indole π-system. In addition, the influence of increased
negative charge at the rear of this pocket agrees with
the previously reported nature of this subsite because
pyridines and derivatives were consistently treated as
neutral species in the 3D-QSAR analyses, although pKa
calculations and measurements suggest a significant
protonation of this center. Hence, this red contour region
highlights an enhanced tendency for a protonated form
to increase biological affinity. This could also indicate
favorable interactions with some of the three backbone
carbonyl groups of Ile175, Thr98, and Glu97 pointing
inward toward the S4 pocket. Detailed inspection of the
50 X-ray structure and related complexes with a pyri-
dine at this position also indicate the influence of a

Factor Xa Inhibition
Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9 3307
Figure 8. Contour maps from CoMSIA model E in combination with compound 48 (10 nM). Hydrogens are omitted. (A) Shown
is the steric std*coeff contour map. Green contours (>80% contribution) refer to regions where additional steric bulk is favorable
for affinity. Yellow contours (<20% contribution) indicate disfavored areas. (B) Shown is the electrostatic std*coeff contour map.
Blue contours (>80% contribution) refer to regions where negatively charged substituents are unfavorable. Red contours (<20%
contribution) indicate regions where negatively charged substituents are favorable. (C) Shown is the hydrophobic std*coeff contour
map. White contours (>80% contribution) refer to regions where hydrophilic substituents are favorable. Yellow contours (<20%
contribution) indicate regions where hydrophobic substituents are favorable. (D) Same as (A) with factor Xa solvent-accessible
surface to illustrate complementarity to protease topology. (E) Same as (B) with factor Xa solvent-accessible surface. (F) Same as
(C) with factor Xa solvent-accessible surface.
structurally conserved water in this position. This water
molecule is within hydrogen-bonding distance to Ile175-
CO (2.62 Å) and Thr89-CO (2.75 Å) and thus only offers
the possibility to interact as acceptor with the nearby
pyridine nitrogen (3.36 Å), which itself is also within a
3.12 Å distance to the Thr98-CO. Maintaining this
balance of water-mediated hydrogen-bonding interac-
tions is therefore essential for favorable fXa binding
affinity.
Many published X-ray structures underscore the
presence of cation-π interactions22,23,66-68 between the
ligand and the Trp215 indole ring in S4. However, it is
possible to compensate for this effect by purely hydro-
phobic and van der Waals interactions with aromatic
side chains of Phe174, Trp215, and Tyr99 surrounding
the S4 pocket19 and positively charged residues that are
in favorable electrostatic interactions with the highly
flexible side chain of Glu97 at the edge of S4. Thus, the
detailed analysis of experimental binding modes has
revealed three subregions as key ligand binding do-
mains within S4: the “hydrophobic box” formed by the
three aromatic amino acids forming a deep aryl-binding
pocket, the “cationic hole” close to Glu97, and the water-
binding site at the rear of S4.
subsite, where protein carbonyl groups point inward
forming a hydrogen-bonding network and the piperidyl
nitrogen is positioned on top of the Trp215 indole ring.
Preferred hydrophobic interactions are indicated by
yellow contour regions (<20% contribution). These
interactions are located in positions 4 and 6 of the
3-oxybenzamide scaffold, in particular on top of the
hydrophobic portion of the Cys192-Cys220 disulfide
bridge within the ester-binding pocket. In addition,
another important contour region is present within the
S1 pocket at the position of the para chlorine substituent
interacting with the aromatic ring of Tyr228, suggesting
the hydrophobic nature of this contact based on the
analysis of this ligand series. From a comparison of fXa
X-ray structures with and without benzamidine in this
pocket, it is known that this position is occupied by a
structurally conserved water in complexes with benz-
amidine derivatives. This water molecule is replaced by
a lipophilic chlorine substituent in the fXa inhibitors
with neutral P1 ligands described here. Hence, this
contour region suggests that the replacement of polar
interactions in the fXa S1 pocket with Asp189 by purely
hydrophobic contacts is possible given the appropriate
size and hydrophobicity of substituents in contact with
the center of the aromatic plane of Tyr228.
The interpretation of additional hydrophobic CoMSIA
fields used for PLS analysis is presented in Figure 8C,F.
White contours indicate regions where hydrophilic
interactions are favorable for affinity (>80% contribu-
tion). Those regions are located at the rear of the S4
4.9. Comparing Tailored Scoring Functions with
fXa Binding Site Topology. The AFMoC std*coeff
fields for model I and six different atom type fields are
shown in Figure 9 with the inhibitor 48 (Ki(fXa) ) 10

3308 Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9
Matter et al.
Figure 9. Contour maps for different field types from AFMoC model I in combination with compound 48 (10 nM). Hydrogens are
omitted. Contours are shown at 75% and 25% contributions for favorable and unfavorable interactions, respectively. (A) C.3:
green contours refer to favored regions, and yellow contours indicate disfavored areas. (B) C.ar: orange contours refer to favored
regions, and white contours indicate disfavored areas. (C) Cl: blue contours refer to favored regions, and red contours indicate
disfavored areas. (D) N.am: blue contours refer to favored regions, and red contours indicate disfavored areas. (E) N.ar: magenta
contours refer to favored regions, and cyan contours indicate disfavored areas. (F) O.3: green contours refer to favored regions,
and yellow contours indicate disfavored areas.
nM). All contours are shown at 75% and 25% contribu-
tions for favorable and unfavorable interactions, respec-
tively. Green contours in Figure 9A indicate regions
where a C.3 atom type in ligands is favorable for fXa
affinity, as exemplified by a contour region situated at
position 4 of the central 3-oxybenzamide. The interpre-
tation of aromatic carbon atom fields (Figure 9B)
indicates favorable contributions (orange) in S1 at the
dichlorophenyl ring, at position 4 of the central 3-oxy-
benzamide and at the distal pyridyl ring situated in S4.
In contrast, the pyridine nitrogen is surrounded by a
white contour region, indicating that replacement by
carbon in this area is detrimental for affinity. Favorable
interactions for chlorine (Figure 9C, red) are present in
the EBP at position 4 of the 3-oxybenzamide and at the
para position of the dichlorophenyl ring in S1, while
unfavorable positions mainly related to nonoptimal fit
of substituents within S1 close to the aromatic plane of
Tyr228 are indicated by blue contours. The interpreta-
tion of the amide nitrogen N.am fields in Figure 9D
positively highlights the 3-oxybenzamide nitrogen as
relevant for affinity, as supported by its involvement
in hydrogen-bonding interactions to fXa, while similar
atom types are unfavorable in the EBP at position 4 of
the 3-oxybenzamide. In Figure 9E the favorable contri-
butions of an aromatic nitrogen N.ar, shown in magenta,
are focused on the pyridine nitrogen atom situated in
S4, while other N.ar substitutions at the distal ring in
S4 or in S1 are detrimental for affinity (cyan). Finally,
the contribution of ether and hydroxyl oxygen O.3 is
indicated by green contours for favorable contributions
at the ether linkage of the oxybenzamide, while yellow
contours in Figure 9F indicate unfavorable positions
mainly in the EBP at position 4 of the benzamide and
at the pyridyl nitrogen in S4 and in S1.
The TScore std*coeff map for model L and amino acid
descriptors are displayed in Figure 10 with 48 as
reference (Ki(fXa) ) 10 nM). Relevant residues for
ChemScore and PLP hydrogen-bonding terms are shown
at levels of -0.04 and +0.04 derived from PLS std*coeff
information (Table 6). ChemScore lipophilic and PLP
steric terms at levels of +0.08 and -0.08 are used to
display favorable and unfavorable amino acid interac-
tions, respectively. Green residues in Figure 10A indi-
cate favorable hydrogen-bonding contributions from the
ChemScore term, in particular for Gly216 in contact
with the 48 3-oxybenzamide nitrogen. Corresponding
results for the PLP hydrogen-bonding term are shown
in Figure 10C. Interactions with positive influence on
affinity are Gln192, limiting the EBP, and Ser214
(green), while negative contributions are associated with
Asp189, Thr98, and Trp215 (yellow). The analysis in
Figure 10B indicates positive contributions from the
ChemScore lipophilic terms in green, while unfavorable
amino acid interactions are indicated in yellow. In
particular, interactions with Ser173 and Cys220 at the
bottom of the EBP are positively correlated to affinity,
while increasing lipophilic interactions with Thr95,
Glu97, Ile175, and Ser195 are detrimental to affinity.
The analysis of the PLP steric contact terms in Figure
10D highlights the importance of amino acids limiting
the EBP and S1 pocket, in particular Arg143, His145,
Lys148, Cys191, Gln192, Gly219, and Cys220 with a
negative contribution (yellow), while interactions of
Glu143, Ser173, and Ser214 are favorable for affinity.
The analysis of the AFMoC ligand-derived and TScore
protein-derived interaction information highlights rel-
evant ligand substructures and protein residues es-
sential for inhibitor binding. Collectively, these models
with the 3D-QSAR analyses from ligands alone might

Factor Xa Inhibition
Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9 3309
Figure 10. Chemical interpretation of TScore model L using the fXa binding site with inhibitor 48. Key amino acids related to
fXa binding affinity are highlighted. Green indicates favorable contributions of a protein-ligand interaction term to affinity based
on PLS derived SD*coeff values, while negative contributions are shown in yellow. (A) SD*coeff map for Chemscore hydrogen
bond term: >0.04 (green) and <-0.04 (yellow). (B) SD*coeff map for Chemscore lipophilic term: >0.08 (green) and <-0.08 (yellow).
(C) SD*coeff map for PLP hydrogen bond term: >0.04 (yellow) and <-0.04 (green; note the inverted sign compared to (A)). (D)
SD*coeff map for PLP steric term: >0.08 (yellow) and <-0.08 (green; note the inverted sign compared to (B)).
be relevant for the design of novel fXa inhibitors further
exploring this interesting chemotype.
difference Glu192/Gln192, which could be targeted by
substituents interacting with the terminal GlnNH2
group in fXa.
4.10. Selectivity toward Thrombin. A set of 21
3-oxybenzamides with acceptable fXa affinity was tested
against the related protease thrombin (Table 1). All
compounds were selective with Ki values larger than 1
µM for thrombin. The best thrombin inhibitor 32 (1097
nM) has a NHMe group directed toward the EBP, which
might favorably interact with Glu192 in thrombin, while
this residue in fXa is Gln192. Replacement of the NHMe
group by a methyl or CF3 in this region led to reduced
thrombin affinity (28, 4866 nM; 29, 31804 nM). This
could be attributed to the lack of favorable hydrogen
bonding to Glu192 or unfavorable electrostatic interac-
tions with this residue. Smaller substituents are toler-
ated, such as chlorine (53, 3064 nM). Adding a second
substituent in the R3 position has a positive effect on
affinity (48, 1801 nM; 49, 1743 nM). Additional steric
bulk in S4, exemplified by 59 with a methyl group ortho
to the distal pyridyl ring, is detrimental for thrombin
affinity (20 777 nM).
The main difference between thrombin and fXa is the
lack of substituents directed toward the thrombin S2
pocket combined with favorable interactions with its
“60”-insertion loop. Different characteristics of the S4
pocket might be important, which preferably binds
hydrophobic substituents in thrombin, while for fXa
basic substituents are accommodated. Additional effects,
as demonstrated in this data set, originate from the
5. Conclusions
While structure-based design aims to understand
relevant protein-ligand interactions, this does not
automatically translate into predictive models for com-
pound prioritization. On the other hand, 3D-QSAR
techniques often cannot be interpreted in protein struc-
tural terms because the derived models do not always
relate to a “bioactive conformation”. In this study we
successfully combined both approaches to unveil ligand
and protein structural requirements for factor Xa
inhibitors in the 3-oxybenzamide series. On the basis
of fXa X-ray structures, consistent and predictive 3D-
QSAR models plus tailored scoring were obtained, which
can be interpreted in terms of important interactions.
These models collectively provided guidelines for ligand
design plus a predictive score for candidates. The
chemical interpretation of CoMFA and CoMSIA contour
maps directly points to those regions in the factor Xa
binding site, where steric, electronic, or hydrophobic
effects play a dominant role in ligand-receptor interac-
tions. The interpretation of 3D-QSAR-derived contour
regions to identify advantageous protein-ligand inter-
actions in combination with the information from X-ray
data allowed the elucidation of main features for the
design of factor Xa inhibitors. The comparable results

3310 Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9
Matter et al.
(16) Rai, R.; Sprengler, P. A.; Elrod, K. C.; Young, W. B. Perspectives
on factor Xa inhibition. Curr. Med. Chem. 2001, 8, 101-119.
(17) Gould, W. R.; Leadley, R. J. Recent advances in the discovery
and development of direct coagulation factor Xa inhibitors. Curr.
Pharm. Des. 2003, 9, 2337-2347.
from 3D-QSAR approaches and their complementarity
to the receptor topology indicate the validity of these
models. Good correlations between predicted and ex-
perimental affinities were obtained from two tailored
scoring functions, namely, AFMoC and TScore. While
3D-QSAR analyses are built on descriptors for the
ligand series alone, those tailored scoring functions use
relevant terms for describing protein-ligand interac-
tions. The quantitative SAR information consistently
extracted from the interpretation of models agrees with
all experimental data for the binding topology and thus
provides guidelines and reasonable activity estimations
for novel inhibitors. From these analyses, a series of
potent, achiral, and selective factor Xa inhibitors emerged
that offer opportunities for compound optimization.
(18) (a) Pinto, D. J. P.; Orwat, M. J.; Wang, S.; Fevig, J. M.; Quan,
M. L.; Amparo, E.; Cacciola, J.; Rossi, K. A.; Alexander, R. S.;
Smallwood, A. M.; Luettgen, J. M.; Liang, L.; Aungst, B. J.;
Wright, M. R.; Knabb, R. M.; Wong, P. C.; Wexler, R. R.; Lam,
P. Y. S. Discovery of 1-[3-(Aminomethyl)phenyl]-N-[3-fluoro-2′-
(methylsulfonyl)-[1,1′-biphenyl]-4-yl]-3-(trifluoromethyl)-1H-pyra-
zole-5-carboxamide (DPC423), a highly potent, selective, and
orally bioavailable inhibitor of blood coagulation factor Xa. J.
Med. Chem. 2001, 44, 566-578. (b) He, W.; Hanney, B.; Myers,
M. R.; Spada, A. R.; Brown, K.; Colussi, D.; Chu, V. Nonbenz-
amidine compounds as selective factor Xa inhibitors. Bioorg.
Med. Chem. Lett. 2000, 15, 1737-1739. (c) Dudley, D. A.;
Bunker, A. M.; Chi, L.; Cody, W. L.; Holland, D. R.; Ignasiak,
D. P.; Janiczek-Dolphin, N.; McClanahan, T. B.; Mertz, T. E.;
Narasimhan, L. S.; Rapundalo, S. T.; Trautschold, J. A.; van
Huis, C. A.; Edmunds, J. J. Rational design, synthesis, and
biological activity of benzoxazinones as novel factor Xa inhibitors.
J. Med. Chem. 2000, 43, 4063-4070. (d) Masters, J. J.; Fran-
ciskovich, J. B.; Tinsley, J. M.; Campbell, C.; Campbell, J. B.;
Craft, T. J.; Froelich, L. L.; Gifford-Moore, D. S.; Hay, L. A.;
Herron, D. K.; Klimkowski, V. J.; Kurz, K. D.; Metz, J. T.; Ratz,
A. M.; Shuman, R. T.; Smith, G. F.; Smith, T.; Towner, R. D.;
Wiley, M. R.; Wilson, A.; Yee, Y. K. Non-amidine-containing 1,2-
dibenzamidobenzene inhibitors of human factor Xa with potent
anticoagulant and antithrombotic activity. J. Med. Chem. 2000,
43, 2087-2092. (e) Herron, D. K.; Goodson, T.; Wiley, M. R.;
Weir, L. C.; Kyle, J. A.; Yee, Y. K.; Tebbe, A. L.; Tinsley, J. M.;
Mendel, D.; Masters, J. J.; Franciskovich, J. B.; Sawyer, J. S.;
Beight, D. W.; Ratz, A. M.; Milot, G.; Hall, S. E.; Klimkowski,
V. J.; Wikel, J. H.; Eastwood, B. J.; Towner, R. D.; Gifford-Moore,
D. S.; Craft, T. J.; Smith, G. F. 1,2-Dibenzamidobenzene inhibi-
tors of human factor Xa. J. Med. Chem. 2000, 43, 859-872. (f)
Wiley, M. R.; Weir, L. C.; Briggs, S.; Bryan, N. A.; Buben, J.;
Campbell, C.; Chirgadze, N. Y.; Conrad, R. C.; Craft, T. J.;
Ficorilli, J. V.; Franciskovich, J. B.; Froelich, L. L.; Gifford-
Moore, D. S.; Goodson, T.; Herron, D. K.; Klimkowski, V. J.;
Kurz, K. D.; Kyle, J. A.; Masters, J. J.; Ratz, A. M.; Milot, G.;
Shuman, R. T.; Smith, T.; Smith, G. F.; Tebbe, A. L.; Tinsley, J.
M.; Towner, R. D.; Wilson, A.; Yee, Y. K. Structure-based design
of potent, amidine-derived inhibitors of factor Xa: Evaluation
of Selectivity, anticoagulant activity, and antithrombotic activity.
J. Med. Chem. 2000, 43, 883-899. (g) Yee, Y. K.; Tebbe, A. L.;
Linebarger, J. H.; Beight, D. W.; Craft, T. J.; Gifford-Moore, D.;
Goodson, T.; Herron, D. K.; Klimkowski, V. J.; Kyle, J. A.;
Sawyer, J. S.; Smith, G. F.; Tinsley, J. M.; Towner, R. D.; Weir,
L.; Wiley, M. R. N2-Aroylanthranilamide inhibitors of human
factor Xa. J. Med. Chem. 2000, 43, 873-882. (h) Choi-Sledeski,
Y. M.; Kearney, R.; Poli, G.; Pauls, H.; Gardner, C.; Gong, Y.;
Becker, M.; Davis, R.; Spada, A.; Liang, G.; Chu, V.; Brown, K.;
Collussi, D.; Leadley, R.; Rebello, S.; Moxey, P.; Morgan, S.;
Bentley, R.; Kasiewski, C.; Maignan, S.; Guilloteau, J.-P.; Mikol,
V. Discovery of an orally efficacious inhibitor of coagulation
factor Xa which incorporates a neutral P1 ligand. J. Med. Chem.
2003, 46, 681-684. (i) Pruitt, J. R.; Pinto, D. J. P.; Galemmo, R.
A.; Alexander, R. S.; Rossi, K. A.; Wells, B. L.; Drummond, S.;
Bostrom, L. L.; Burdick, D.; Bruckner, R.; Chen, H.; Smallwood,
A.; Wong, P. C.; Wright, M. R.; Bai, S.; Luettgen, J. M.; Knabb,
R. M.; Lam, P. Y. S.; Wexler, R. R. Discovery of 1-(2-aminom-
ethylphenyl)-3-trifluoromethyl-N-[3-fluoro-2′-(aminosulfonyl)-
Acknowledgment. We gratefully acknowledge the
programming support of C. Giegerich and the technical
assistance of A. Liesum, V. Brachvogel, P. Loenze, L.
Bayer, S. Schu¨lke, D. Timme, N. Laub, S. Engel, A.
Sihorsch, and M. Ka¨mmerer (Aventis Pharma). We
thank G. Klebe and H. Gohlke (University of Marburg)
for a version of DrugScore including AFMoC. Further-
more, we gratefully acknowledge many stimulating
discussions with F. Al-Obeidi, A. Bauer, J. Czech, O.
Klingler, M. Lorenz, S. Maignan, A. Peyman, K. Ritter,
M. Urmann, M. Wagner, and G. Zoller (Aventis Phar-
ma). The coordinates of the fXa/50 complex were de-
posited in the Brookhaven Protein Data Bank (PDB:
2bmg).
References
(1) Vlasuk, G. P. Direct Factor Xa Inhibitors. In New Therapeutic
Agents in Thrombosis and Thrombolysis; Sasahara, A. A.,
Loscalzo, J., Eds.; Marcel Dekker: New York, 1997; pp 261-
283.
(2) Coleman, R. W., Marder, V. J., Salzman, E. W., Eds. Hemostasis
and Thrombosis: Basic Principles and Clinical Practice; J. B.
Lippincott: Philadelphia, PA, 1994.
(3) Davie, E. W.; Fujikawa, K.; Kisiel, W. The coagulation cascade:
initiation, maintenance, and regulation. Biochemistry 1991, 30,
10363-10370.
(4) Leadley, R. J. Coagulation factor Xa inhibition: biological
background and rationale. Curr. Top. Med. Chem. 2001, 1, 151-
159.
(5) Mann, K. G.; Nesheim, M. E.; Church, W. R.; Haley, P.;
Krishnaswamy, S. Surface dependent reactions of the vitamin
K-dependent enzyme complexes. Blood 1990, 76, 1-16.
(6) Ahmad, S. S.; Rawala-Sheikh, R.; Walsh, P. N. Compounds and
assembly of the factor Xa activating complex. Semin. Thromb.
Hemostasis 1992, 18, 311-323.
(7) Harker, L. A.; Hanson, S. R.; Kelly, A. B. Antithrombotic
strategies targeting thrombin activities, thrombin receptors and
thrombin generation. Thromb. Haemostasis 1997, 78, 736-741.
(8) Waxman, L.; Smith, D. E.; Arcuri, K. E.; Vlasuk, G. P. Tick
[1,1′-biphenyl)]-4-yl]-1H-pyrazole-5-carboxamide (DPC602),
a
potent, selective, and orally bioavailable factor Xa inhibitor. J.
Med. Chem. 2003, 46, 5298-5315.
anticoagulant peptide (TAP) is
a novel inhibitor of blood
coagulation factor Xa. Science 1990, 248, 593-596.
(9) Hara, T.; Yokoyoma, A.; Tanabe, K.; Ishihara, H.; Iwamoto, M.
DX-9065a, an orally active, specific inhibitor of factor Xa, inhibits
thrombosis without affecting bleeding time in rats. Thromb.
Haemostasis 1995, 74, 635-639.
(19) Matter, H.; Defossa, E.; Heinelt, U.; Blohm, P.-M.; Schneider,
D.; Mueller, A.; Herok, S.; Schreuder, H.; Liesum, A.; Brachvogel,
V.; Loenze, P.; Walser, A.; Al-Obeidi, F.; Wildgoose, P. Design
and quantitative structure-activity relationship of 3-amidi-
nobenzyl-1H-indole-2-carboxamides as potent, nonchiral, and
selective inhibitors of blood coagulation factor Xa. J. Med. Chem.
2002, 45, 2749-2769.
(20) Maignan, S.; Mikol, V. The use of 3D structural data in the
design of specific factor Xa inhibitors. Curr. Top. Med. Chem.
2001, 1, 161-174.
(21) Nazare´, M.; Matter, H.; Klingler, O.; Al-Obeidi, F.; Schreuder,
H.; Zoller, G.; Czech, J.; Lorenz, M.; Dudda, A.; Peyman, A.;
Nestler, H. P.; Urmann, M.; Bauer, A.; Laux, V.; Wehner, V.;
Will, D. W. Novel factor Xa inhibitors based on a benzoic acid
scaffold and incorporating a neutral P1 ligand. Bioorg. Med.
Chem. Lett. 2004, 14, 2801-2805.
(22) Padmanabhan, K.; Padmanabhan, K. P.; Tulinsky, A.; Park, C.
H.; Bode, W.; Huber, R.; Blankenship, D. T.; Cardin, A. D.; Kisiel,
W. Structure of human Des(1-45) factor Xa at 2.2 Å resolution.
J. Mol. Biol. 1993, 232, 947-966.
(10) Antman, E. M. Hirudin in acute myocardial infarction. Safety
report from the thrombolysis and thrombin inhibition in myo-
cardial infarction (TIMI) 9A trial. Circulation 1994, 90, 1624-
1630.
(11) Philippides, G. J.; Loscalzo, J. Potential advantages of direct-
acting thrombin inhibitors. Coron. Artery Dis. 1996, 7, 497-507.
(12) Al-Obeidi, F.; Ostrem, J. A. Factor Xa inhibitors. Expert Opin.
Ther. Pat. 1999, 74, 635-639.
(13) Ewing, W. R.; Pauls, H. W.; Spada, A. P. Progress in the design
of inhibitors of coagulation factor Xa. Drugs Future 1999, 24,
771-787.
(14) Zhu, B.-Y.; Scarborough, R. M. Factor Xa inhibitors: recent
advances in anticoagulant agents. Annu. Rep. Med. Chem. 2000,
35, 83-102.
(15) Fevig, J. M.; Wexler, R. R. Thrombin and factor Xa inhibitors.
Annu. Rep. Med. Chem. 1999, 34, 81-100.

Factor Xa Inhibition
Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9 3311
(23) Maignan, S.; Guilloteau, J.-P.; Choi-Sledeski, Y. M.; Becker, M.
R.; Ewing, W. R.; Pauls, H. W.; Spada, A. P.; Mikol, V. Molecular
structures of human factor Xa complexed with ketopiperazine
inhibitors: Preference for a neutral group in the S1 pocket. J.
Med. Chem. 2003, 46, 685-690.
(24) Adler, M.; Kochanny, M. J.; Ye, B.; Rumnenik, G.; Light, D. R.;
Biancalana, S.; Whitlow, M. Crystal Structures of two potent
nonamidine inhibitors bound to factor Xa. Biochemistry 2002,
41, 15514-15523.
(39) Goodford, P. J. A computational procedure for determining
energetically favorable binding sites on biologically important
macromolecules. J. Med. Chem. 1985, 28, 849-857.
(40) McMartin, C.; Bohacek, R. S. QXP: powerful, rapid computer
algorithms for structure-based drug design. J. Comput.-Aided
Mol. Des. 1997, 11, 333-344.
(41) Weiner, S. J.; Kollman, P. A.; Case, D. A.; Singh, U. C.; Ghio,
C.; Algona, G.; Profeta, S.; Weiner, P. A new force field for
molecular mechanical simulation of nucleic acids and proteins.
J. Am. Chem. Soc. 1984, 106, 765-784.
(25) (a) Maignan, S.; Guilloteau, J.-P.; Pouzieux, S.; Choi-Sledeski,
Y. M.; Becker, M. R.; Klein, S. I.; Ewing, W. R.; Pauls, H. W.;
Spada, A. P.; Mikol, V. Crystal structures of human factor Xa
complexed with potent inhibitors. J. Med. Chem. 2000, 43,
3226-3232. (b) Kamata, K.; Kawamoto, H.; Honma, T.; Iwama,
T.; Kim, S.-H. Structural basis for chemical inhibition of human
blood coagulation factor Xa. Proc. Natl. Acad. Sci. U.S.A. 1998,
95, 6630-6635. (c) Brandstetter, H.; Ku¨hne, A.; Bode, W.; Huber,
R.; von der Saal, W.; Withensohn, K.; Engh, R. A. X-ray structure
of active site-inhibited clotting factor Xa. Implications for drug
design and substrate recognition. J. Biol. Chem. 1996, 271,
29988-29992. (d) Nar, H.; Bauer, M.; Schmid, A.; Stassen, J.-
M.; Wienen, W.; Priepke, H. W. M.; Kauffmann, I. K.; Ries, U.
J.; Hauel, N. H. Structural basis for inhibition promiscuity of
dual specific thrombin and factor Xa blood coagulation inhibitors.
Structure 2001, 9, 29-37. (e) Adler, M.; Davey, D. D.; Phillips,
G. B.; Kim, S.-H.; Jancarik, J.; Rumennik, G.; Light, D. R.;
Whitlow, M. Preparation, characterization and the crystal
structure of the inhibitor ZK-807834 (CI-1031) complexed with
factor Xa. Biochemistry 2000, 39, 12534-12542.
(26) Cramer, R. D.; Patterson, D. E.; Bunce, J. E. Comparative
molecular field analysis (CoMFA). 1. Effect of shape on binding
of steroids to carrier proteins. J. Am. Chem. Soc. 1988, 110,
5959-5967.
(27) Clark, M.; Cramer, R. D.; Jones, D. M.; Patterson, D. E.;
Simeroth, P. E. Comparative molecular field analysis (CoMFA).
2. Towards its use with 3D-structural databases. Tetrahedron
Comput. Methodol. 1990, 3, 47-59.
(42) (a) Halgren, T. Maximally diagonal force constants in dependent
angle-bending coordinates. II. Implications for the design of
empirical force fields. J. Am. Chem. Soc. 1990, 112, 4710-4723.
(b) Halgren, T. MMFF VI. MMFF94s option for energy minimi-
zation studies. J. Comput. Chem. 1999, 20, 720-729.
(43) SYBYL Molecular Modelling Package, version 6.9; Tripos: St.
Louis, MO, 2003.
(44) (a) Heiden, W.; Goetze, T.; Brickmann, J. Fast generation of
molecular surfaces from 3D data fields with enhanced “marching
cube” algorithm. J. Comput. Chem. 1993, 14, 246-250. (b)
Brickmann, J.; Bertling, H.; Bussian, B. M.; Goetze, T.; Kno-
blauch, M.; Waldherr-Teschner, M. MOLCADsinteractive mo-
lecular computer graphics on high-performance computers.
Tagungsber.sVortragstag., Ges. Dtsch. Chem., Fachgruppe Chem.-
Inf. 1987, 93-111.
(45) Heiden, W.; Moeckel, G.; Brickmann, J. A new approach to the
display of local lipophilicity/hydrophilicity mapped on molecular
surfaces. J. Comput.-Aided Mol. Des. 1993, 7, 503-514.
(46) (a) Ghose, A.; Crippen, G. Atomic physicochemical parameters
for three-dimensional structure-directed quantitative structure-
activity relationships. 1. Partition coefficients as a measure of
hydrophobicity. J. Comput. Chem. 1986, 7, 565-577. (b) Viswa-
nadhan, V. N.; Ghose, A. K.; Revankar, G. R.; Robins, R. K.
Atomic physicochemical parameters for three-dimensional struc-
ture-directed quantitative structure-activity relationships. 4.
Additional parameters for hydrophobic and dispersive interac-
tions and their application for an automated superposition of
certain naturally occurring nucleoside antibiotics. J. Chem. Inf.
Comput. Sci. 1989, 29, 163-172.
(47) (a) Connolly, M. L. Solvent-accessible surfaces of proteins and
nucleic acids. Science 1983, 221, 709-713. (b) Connolly, M. L.
Analytical molecular surface calculation. J. Appl. Crystallogr.
1983, 16, 548-558.
(48) (a) Wold, S. Cross-validatory estimation of the number of
component in factor and principal component models. Techno-
metrics 1978, 4, 397-405. (b) Diaconis, P.; Efron, B. Computer-
intensive methods for statistics. Sci. Am. 1984, 116, 96-117.
(c) Cramer, R. D.; Bunce, J. D.; Patterson, D. E. Crossvalidation,
bootstrapping and partial least squares compared with multiple
regression in conventional QSAR studies. Quant, Struct.-Act.
Relat. 1988, 7, 18-25.
(28) 3D-QSAR in Drug Design. Theory, Methods and Applications;
Kubinyi, H., Ed.; ESCOM: Leiden, The Netherlands, 1993.
(29) Klebe, G.; Abraham, U.; Mietzner, T. Molecular similarity indices
in a comparative analysis (CoMSIA) of drug molecules to
correlate and predict their biological activity. J. Med. Chem.
1994, 37, 4130-4146.
(30) Gohlke, H.; Klebe, G. DrugScore meets CoMFA: Adaption of
fields for molecular comparison (AFMoC) or how to tailor
knowledge-based pair potentials to a particular protein. J. Med.
Chem. 2002, 45, 4153-4170.
(31) Matter, H.; Giegerich, C. Unpublished results.
(32) Tucker, T. J.; Brady, S. F.; Lumma, W. C.; Lewis, S. D.; Gardell,
S. J.; Naylor-Olsen, A. M.; Yan, Y.; Sisko, J. T.; Stauffer, K. J.;
Lucas, B. J.; Lynch, J. J.; Cook, J. J.; Stranieri, M. T.; Holahan,
M. A.; Lyle, E. A.; Baskin, E. P.; Chen, I. W.; Dancheck, K. B.;
Krueger, J. A.; Cooper, C. M.; Vacca, J. P. Design and synthesis
of a series of potent and orally bioavailable noncovalent thrombin
inhibitors that utilize nonbasic groups in the P1 position. J. Med.
Chem. 1998, 41, 3210-3219.
(49) Sheridan, R. P.; Nachbar, R. B.; Bush, B. L. Extending the trend
vector: The trend matrix and sample-based partial least squares.
J. Comput.-Aided Mol. Des. 1994, 8, 323-340.
(50) (a) Wold, S.; Albano, C.; Dunn, W. J.; Edlund, U.; Esbenson, K.;
Geladi, P.; Hellberg, S.; Lindberg, W.; Sjo¨stro¨m, M. In Chemo-
metrics: Mathematics and Statistics in Chemistry; Kowalski, B.,
Ed.; Reidel: Dortrecht, The Netherlands, 1984; pp 17-95. (b)
Dunn, W. J.; Wold, S.; Edlund, U.; Hellberg, S.; Gasteiger, J.
Multivariate structure-activity relationship between data from
a battery of biological tests and an ensemble of structure
descriptors: The PLS method. Quant. Struct.-Act. Relat. 1984,
3, 31-137. (c) Geladi, P. Notes on the history and nature of
partial least squares (PLS) modelling. J. Chemom. 1988, 2, 231-
246.
(33) Nazare, M.; Will, D. W.; Peyman, A.; Matter, H.; Zoller, G.;
Gerlach, U. Preparation of heterocyclyl benzamides as inhibitors
of factor Xa and factor VIIa. Eur. Pat. Appl. EP1217000, 2002.
(34) (a) Mitsunobu, O. The use of diethyl azodicarboxylate and
triphenylphosphine in synthesis and transformation of natural
products. Synthesis 1981, 1, 1-28. (b) Tunoori, A. R.; Dutta, D.;
Georg, G. I. Polymer-bound triphenylphosphine as traceless
reagent for Mitsunobu reactions in combinatorial chemistry:
synthesis of aryl ethers from phenols and alcohols. Tetrahedron
Lett. 1998, 39, 8751-8754.
(35) Ostrem, J. A.; Al-Obeidi, F. A.; Safar, P.; Safarova, A.; Stringer,
S. K.; Patek, M.; Cross, M. T.; Spoonamore, J.; LoCascio, J. C.;
Kasireddy, P.; Thorpe, D.; Sepetov, N.; Lebl, M.; Wildgoose, P.;
Strop, P. Discovery of a novel, potent, and specific family of factor
Xa inhibitors via combinatorial chemistry. Biochemistry 1998,
37, 1053-1059.
(51) Clark, M.; Cramer, R. D. The probability of chance correlation
using partial least squares (PLS). Quant. Struct.-Act. Relat.
1993, 12, 137-145.
(52) (a) Leonard, J. M.; Clark, R. D. PLS validation by progressive
scrambling. Presented at the 12th European Symposium on
Quantitative Structure-Activity Relationships (QSAR’98), 1998.
(b) Clark, R. D.; Fox, P. C. Statistical variation in progressive
scrambling. J. Comput.-Aided Mol. Des. 2004, 18, 563-576.
(53) Matter, H.; Kotsonis, P.; Klingler, O.; Strobel, H.; Fro¨hlich, L.;
Pfleiderer, W.; Schmidt, H. H. H. W. Structural requirements
for inhibition of the neuronal nitric oxide synthase (NOS-I): 3D-
QSAR analysis of 4-oxo and 4-amino-pteridine based antago-
nists. J. Med. Chem. 2002, 45, 2923-2941.
(36) Kabsch, W. Evaluation of single-crystal X-ray diffraction data
from a position sensitive detector. J. Appl. Crystallogr. 1988,
21, 916-924.
(37) Bru¨nger, A. T. X-PLOR: A System for X-ray Crystallography
and NMR, version 3.1; Yale University Press: New Haven, CT,
1992.
(54) UNITY Chemical Information Software, version 4.4; Tripos Inc.
(1699 S. Hanley Road, St. Louis, MO 63144).
(38) PDB files from RSCB Protein Database: http://www.rcsb.org/
pdb. (a) Bernstein, F. C.; Koetzle, T. F.; Williams, G. J. B.; Meyer,
E. F.; Brice, M. D.; Rodgers, J. R.; Kennard, O.; Shimanouchi,
T.; Tasumi, M. The Protein Data Bank: a computer-based
archival file for macromolecular structures. J. Mol. Biol. 1977,
112, 535-542. (b) Berman, J. W.; Zukang, F.; Gilliland, G.; Bhat,
T. N.; Weissig, H.; Shindyalov, I. N.; Bourne, P. E. The Protein
Data Bank. Nucleic Acids Res. 2000, 28, 235-242.
(55) (a) Willett, P. Similarity and Clustering in Chemical Information
Systems; Research Studies Press: Letchworth, England, 1987.
(b) Willett, P.; Winterman, V. A. Comparison of some measures
for the determination of inter-molecular structural similarity.
Quant. Struct.-Act. Relat. 1986, 5, 18-25.
(56) Gohlke, H.; Hendlich, M.; Klebe, G. Knowledge-based scoring
function to prodict protein-ligand interactions. J. Mol. Biol.
2000, 295, 337-356.

3312 Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9
Matter et al.
(57) Gohlke, H.; Hendlich, M.; Klebe, G. Predicting binding modes,
binding affinities and “hot spots” for protein-ligand complexes
using a knowledge-based scoring function. Perspect. Drug Dis-
covery Des. 2000, 20, 115-144.
(64) Po¨tter, T.; Matter, H. Random or rational design? Evaluation of
diverse compound subsets from chemical structure databases.
J. Med. Chem. 1998, 41, 478-488.
(65) (a) Klebe, G.; Abraham, U. Comparative molecular similarity
index analysis (CoMSIA) to study hydrogen-bonding properties
and to score combinatorial libraries. J. Comput.-Aided Mol. Des.
1999, 13, 1-10. (b) Bo¨hm, M.; Stu¨rzebecher, J.; Klebe, G. Three-
dimensional quantitative structure-activity relationship analy-
ses using comparative molecular field analysis and comparative
molecular similarity indices analysis to elucidate selectivity
differences of inhibitors binding to trypsin, thrombin, and factor
Xa. J. Med. Chem. 1999, 42, 458-477. (c) Klebe, G. Comparative
molecular similarity indices analysis. CoMSIA. Perspect. Drug
Discovery Des. 1998, 12/13/14, 87-104.
(66) (a) Gallivan, J. P.; Dougherty, D. A. A computational study of
cation-π interactions vs Salt bridges in aqueous media: impli-
cations for protein engineering. J. Am. Chem. Soc. 2000, 122,
870-874. (b) Gallivan, J. P.; Dougherty, D. A. Cation-π interac-
tions in structural biology. Proc. Natl. Acad. Sci. U.S.A. 1999,
96, 9459-9464. (c) Ma, J. C.; Dougherty, D. A. The cation-π
interaction. Chem. Rev. 1997, 97, 1303-1324.
(58) Denoted as LOO_RSQ and PLS_RSQ in AFMoC from DrugScore
1.2.
(59) Eldridge, M. D.; Murray, C. W.; Auton, T. R.; Paolini, G. V.; Mee,
R. P. Empirical scoring functions: I. The development of a fast
empirical function to estimate the binding affinity of ligands in
receptor complexes. J. Comput.-Aided Mol. Des. 1997, 47, 425-
445.
(60) Gehlhaar, D. K.; Verkhivker; Rejto, P. A.; Sherman, C. J.; Fogel,
D. B.; Fogel, L. J.; Freer, S. T. Molecular recognition of the
inhibitor AG-1343 by HIV-1 protease: conformationally flexible
docking by evolutionary programming. Chem. Biol. 1995, 2,
317-324.
(61) (a) ReliBase, version 1.2.3; Cambridge Crystallographic Data
Centre (CCDC) (12 Union Road, Cambridge CB2 1EZ, U.K.),
January 2004. (b) Hendlich, M.; Bergner, A.; Gu¨nther, J.; Klebe,
G. Relibase: design and development of a database for compre-
hensive analysis of protein-ligand interactions. J. Mol. Biol.
2003, 326, 607-620. (c) Gu¨nther, J.; Bergner, A.; Hendlich, M.;
Klebe, G. Utilising structural knowledge in drug design strate-
gies: Applications using relibase. J. Mol. Biol. 2003, 326, 621-
636.
(67) Flocco, M. M.; Mowbray, S. L. Planar stacking interactions of
arginine and aromatic side-chains in proteins. J. Mol. Biol. 1994,
235, 709-717.
(68) Sulpizi, M.; Carloni, P. Cation-π versus OH-π interactions in
proteins: A density functional study. J. Chem. Phys. B 2000,
104, 10087-10091.
(62) Dunitz, J. D. The cost of bound water in crystals and biomol-
ecules. Science 1994, 264, 670-671.
(63) Verdonk, M. L.; Cole, J. C.; Watson, P.; Gillet, V.; Willett, P.
SuperStar: Improved knowledge-based interaction fields for
protein binding sites. J. Mol. Biol. 2001, 307, 841-859.
JM049187L