Conformer generation with OMEGA: algorithm and validation using high quality structures from the Protein Databank and Cambridge Structural Database.

Abstract

Here, we present the algorithm and validation for OMEGA, a systematic, knowledge-based conformer generator. The algorithm consists of three phases: assembly of an initial 3D structure from a library of fragments; exhaustive enumeration of all rotatable torsions using values drawn from a knowledge-based list of angles, thereby generating a large set of conformations; and sampling of this set by geometric and energy criteria. Validation of conformer generators like OMEGA has often been undertaken by comparing computed conformer sets to experimental molecular conformations from crystallography, usually from the Protein Databank (PDB). Such an approach is fraught with difficulty due to the systematic problems with small molecule structures in the PDB. Methods are presented to identify a diverse set of small molecule structures from cocomplexes in the PDB that has maximal reliability. A challenging set of 197 high quality, carefully selected ligand structures from well-solved models was obtained using these methods. This set will provide a sound basis for comparison and validation of conformer generators in the future. Validation results from this set are compared to the results using structures of a set of druglike molecules extracted from the Cambridge Structural Database (CSD). OMEGA is found to perform very well in reproducing the crystallographic conformations from both these data sets using two complementary metrics of success.

Full text

Page 1

Conformer Generation with OMEGA: Algorithm and Validation Using High Quality
Structures from the Protein Databank and Cambridge Structural Database
Paul C. D. Hawkins,* A. Geoffrey Skillman, Gregory L. Warren, Benjamin A. Ellingson, and
Matthew T. Stahl
OpenEye Scientific Software, 9 Bisbee Court, Suite D, Santa Fe, New Mexico 87508
Received January 21, 2010
Here, we present the algorithm and validation for OMEGA, a systematic, knowledge-based conformer
generator. The algorithm consists of three phases: assembly of an initial 3D structure from a library of
fragments; exhaustive enumeration of all rotatable torsions using values drawn from a knowledge-based list
of angles, thereby generating a large set of conformations; and sampling of this set by geometric and energy
criteria. Validation of conformer generators like OMEGA has often been undertaken by comparing computed
conformer sets to experimental molecular conformations from crystallography, usually from the Protein
Databank (PDB). Such an approach is fraught with difficulty due to the systematic problems with small
molecule structures in the PDB. Methods are presented to identify a diverse set of small molecule structures
from cocomplexes in the PDB that has maximal reliability. A challenging set of 197 high quality, carefully
selected ligand structures from well-solved models was obtained using these methods. This set will provide
a sound basis for comparison and validation of conformer generators in the future. Validation results from
this set are compared to the results using structures of a set of druglike molecules extracted from the Cambridge
Structural Database (CSD). OMEGA is found to perform very well in reproducing the crystallographic
conformations from both these data sets using two complementary metrics of success.
INTRODUCTION
Conformer generation has been a topic of considerable
interest to the modeling community for a number of years,
and a large number of methods for conformer generation
have been presented.1-8These methods have utilized a wide
variety of approaches to the problem, including systematic
enumeration, knowledge-based rule sets, random coordinate
embedding, and energy-based or energy-biased sampling. All
these approaches attempt to generate a set of conformations
designated to be different from one another by some measure,
usually geometric, and thereby aim to sample some subset
of the conformational space available to a molecule. The
method used by OMEGA is based on random coordinate
embedding and refinement, followed by torsion driving using
a set of rules derived from experimental structures and energy
profile calculations, and then subsequent sampling of the
resulting conformer ensemble by geometric and energetic
criteria.
Publications that demonstrate the utility of conformer
generators have approached the problem of validation of such
tools in a number of different ways. Conformer generation
is not an end in itself but a means to obtain input for other
applications, most frequently shape comparison tools (or
other 3D ligand matching approaches) and docking engines.
Indeed conformer generation is often subsumed as a portion
of a docking algorithm, which prevents independent evalu-
ation of the conformer generation within such tools. One
approach to validation has, therefore, been to generate
conformer ensembles for use in a downstream tool (e.g.,
pharmacophore perception).9The quality of the conformer
ensembles is then judged by the quality of the output from
the downstream tool. While it might be argued that the “real-
life” use of conformer generators is well reflected in this
approach, judging the quality of a conformer generator based
on the performance of a separate tool is fraught with
problems. If this tool has been optimized to use conformers
from a certain engine, then it would be expected to produce
“better” results from this engine over all others. Other
approaches to validating conformer generators have empha-
sized estimation of conformational coverage based on both
pharmacophoric and geometric measures.10In studies of this
kind, coverage of conformational space is held to be an
absolute good and the tools under study are assessed based
on their ability to cover the greatest amount of conforma-
tional space. However, there are large areas of “accessible”
conformational space for a flexible molecule that are
populated by conformers very different in shape and
geometry from any experimentally determined conformer (be
it in the solid state, in solution, or in the gas phase). As such,
simply maximizing coverage of conformational space may
result in generation of a conformer set containing a high
proportion of conformers irrelevant to any measurable
experimental property of the molecule.
By far, the most common, though not always the most
well-executed, method for validation of conformer genera-
tors, and the method used in this paper, is the reproduction
of experimental crystal structures11(often those retrieved
from the Protein Databank (PDB)12). The implicit argument
proffered for this approach is that coverage of conformational
space near an experimental conformation of a molecule is
more important than coverage in other regions of confor-
* Corresponding author e-mail: phawkins@eyesopen.com; tel.: 505-473-
7385ext. 65.
J. Chem. Inf. Model. 2010, 50, 572–584
572
10.1021/ci100031x
 2010 American Chemical Society
Published on Web 03/17/2010

Page 2

mational space. Since experimental structures, while fre-
quently not at a local minimum for a force field,13are
generally considered to exhibit low levels of strain,14the
regions of conformational space near an experimental
structure can be considered as basins in the molecule’s energy
profile. Accordingly, assessing sampling in these regions of
the energy hypersurface can be a good guide to the
performance of a method in generating low-energy confor-
mations. To perform this sort of validation, a number of
protein-ligand cocomplexes are selected and the ligands are
extracted. The connection tables of these ligands are used
as input for the conformer generator to provide a set of
conformers for each ligand, unbiased by any structural
information of the original experimental conformation. A
metric of quality is then calculated that provides a measure
of the three-dimensional similarity between the experimental
conformation for the ligand and the conformers in the set.
There are, therefore, two critical choices to this method of
validation: selection of the appropriate metric of quality to
perform the comparison and selection of suitable cocrystal
structures to provide the ligands. While many metrics for
comparing computed and experimental conformations are
available, such as rms error in Cartesian space (rmsd), rms
error in torsion space (RMST), and relative displacement
error (RDE),15choosing the metric of quality presents some
subtle problems. Atom-based metrics like rmsd, while very
popular, have significant problems in their interpretation and
necessarily have a dependence for their significance on the
experimental error/precision of the crystal structure (a
particularly difficult issue for PDB ligands). In addition, RDE
and rmsd have no easily calculable upper bound (unlike
RMST), and all these measures have a size dependence that
is not easily compensated for. In this paper, two different
metrics are used to compare computed and experimental
conformations of druglike molecules, one atom based (rmsd)
and one shape based (Tanimoto combo). The conformational
coverage close to the experimental conformation is also
assessed. In this way, it is hoped to provide a more complete
picture of the ability of OMEGA (or any other conformer
generator) to match experimental conformations. Identifying
suitable cocomplex structures from the PDB to provide the
experimental conformations of small molecules is a difficult
and complex process. Unfortunately, in most publications
in this area, insufficient care is taken in the selection of the
structures, resulting in unwarranted or questionable conclu-
sions being drawn based on unreliable data. Some of the
issues that should be considered in the selection of structures
from the PDB include the following:
Are the deposited models of good quality at both the global
and local levels? There is little sense in trying to reproduce
a deposited structure that is a poor model of the experimental
density or a model that is not completely defined by the
density.
Do the structures possess an appropriate degree of preci-
sion in their atomic coordinates? When using atom-based
metrics like rmsd, paying attention to the precision in the
atomic coordinates is critical: an rmsd or any measure of
conformer reproduction more precise than the inherent
inaccuracy/experimental error in the atomic coordinates is
meaningless. Do the structures show intra- or intermolecular
clashes? Significant atomic clashes, either between the ligand
and the protein or between atoms of the ligand, must be
artifacts of the process of fitting the ligand to the experi-
mental electron density and not an inherent property of the
structure. The binding energy available upon protein-ligand
complexation is unlikely to be sufficient to drive atoms into
close contact with one another. Are the ligands covering a
reasonable amount of structural or chemical space? A set of
ligands with a large degree of structural redundancy does
not provide as rigorous a test of conformer generators as a
structurally nonredundant set does. Also, in order for the
results of the validation to be useful in a predictive or
prospective manner, they should be readily generalizable to
any set of applicable molecules. This can only be done with
validation sets that cover chemical space in a reasonable way.
Consideration of these questions has allowed the develop-
ment of quantitative criteria for structural quality in the PDB.
These criteria were applied as successive filters to three
previously published large data sets of cocrystal structures,
PDBbind,16the Sadowski set13and the Kirchmair set,17to
provide a high quality group of cocrystal structures to es-
tablish a gold standard for the validation of approaches to
conformer sampling. A set of small molecule structures from
the Cambridge Structural Database (CSD)18was also col-
lected to provide complementary data to that obtained with
the ligands from the PDB. These two data sets are applied
to the evaluation of the OMEGA conformer generation
application using the metrics mentioned above.
METHODS
Here, the OMEGA algorithm is introduced in some detail;
the metrics for comparison of the generated conformers to
the experimental structures are outlined, and the criteria for
the identification of suitable protein-ligand cocomplexes and
small molecule structures with which to validate the algo-
rithm are discussed.
OMEGA Algorithm. The design goal for OMEGA is to
provide a thorough, though not exhaustive, sampling of
conformational space for druglike molecules at as high a
speed as possible (typically 2-2.5 s per molecule on a
workstation running SLED 10 with a 2.4 GHz processor and
4 GB of RAM). The process used by OMEGA for conformer
generation can be considered in several parts, two of which
are generic and are precalculated and the rest of which are
molecule specific and occur at run time. The first two steps
are construction of a database of fragments from which the
molecule will be assembled and derivation of a torsion
library; the remaining steps generate a large ensemble of
conformations and sample from this ensemble to deliver the
final set of conformers.
1. Fragment Database Preparation. The OMEGA frag-
ment database contains one or more 3D conformations for
every entry (flexible rings have multiple conformations, rigid
rings and acyclic groups have one conformation). The
database is generated by fragmenting a very large collection
of commercially available compounds into contiguous ring
systems and small linear linkers. One or more 3D conforma-
tions for each fragment are generated by the following
procedure: A distance bounds matrix is generated based on
the connection table of the fragment; the distance bounds
are augmented by volume constraints for chirality and
planarity; the coordinates of each atom are randomly
embedded in a Cartesian space and optimized to fulfill the
CONFORMER GENERATION WITH OMEGA
J. Chem. Inf. Model., Vol. 50, No. 4, 2010 573

Page 3

bounds and constraints; the fragment is further refined against
a modified version of the Merck molecular force field
(MMFF94),19in which the electrostatic and attractive van
der Waals’ terms are removed. The procedure is repeated
hundreds of times for each fragment, with cycles being
sampled more heavily than acyclics, and the lowest energy
conformers that are unique in rmsd space are retained.
2. Torsion Sampling Dictionary. A hierarchy of torsions
defined by SMARTS patterns is preorganized so that every
rotatable bond in a small molecule matches at least one
torsion rule. Every torsion definition is associated with a list
of angles at which the torsion should be sampled. Very
common or structurally unique torsions may be represented
by very specific rules, while unusual torsions are only
covered by generic torsion rules, often with relatively heavy
sampling. Specific rules are employed when the barrier to
rotation is high and the minima in the torsion profile are
well-known. The torsion angles specified by the torsion
library were derived by analysis of a set of experimental
crystal structures (mostly from the PDB) and from energy
scans of certain torsions against MMFF94.
3. 3D Structure Generation. OMEGA requires a com-
pletely connected molecular graph as input. Each graph is
fragmented in the same manner as the fragment database
and fragment conformations are drawn directly from the
fragment database described in step 1. If a molecular
fragment is not found in the fragment database, the procedure
described in step 1 is carried out on-the-fly to generate
coordinates for the missing fragment. The fragments are
assembled into the parent molecule by overlapping fragments
using geometric and chemical rules, thereby providing one
or a small number of initial conformations for the molecule.
This set of conformers includes all necessary ring conformer
sampling.
4. Torsion DriVing. Every rotatable bond in the conform-
ers generated in step 3 is compared to the torsion library
from step 2, and the appropriate torsion angles are noted. A
torsion buildup procedure is applied to all the torsions in
the molecule to generate a large ensemble of conformations
that does not contain severe internal clashes or duplicates
due to common symmetries.
5. Sampling. The design goal of OMEGA is to produce
up to a few hundred low energy, structurally diverse
conformers, and as such, the much larger collection of
conformers produced by step 4 must be sampled. This
process is begun by ordering the conformers using a simple
scoring function that eliminates conformers with internal
clashes. While many functions are suitable for this process,
OMEGA uses a modified form of MMFF94 (vide supra).
Next, beginning with the lowest scoring conformer, all higher
scoring conformers that are less than a user-defined rmsd
(including heavy atom automorphisms) to the lower scoring
conformer are eliminated (the default rmsd cutoff is 0.5 Å).
This process is continued with each successive conformer
in the ordered list until (a) all conformers have a score less
than 10 units higher than the lowest energy conformer
produced or (b) 200 mutually unique conformers are identi-
fied. At no stage in this process is minimization performed
because this tends to produce highly compact (folded)
conformations not reflective of the conformation(s) found
in solution or bound to a protein. This is mostly due to the
differences between the gas phase and the solution phase or
complexed potential energy surfaces for small molecules.
Extensive experimentation in-house has shown that mini-
mization of ensembles of conformers either before or after
deduplication improves neither solid-state structure reproduc-
tion nor virtual screening performance with tools like ROCS
or FRED.
Suitable Metrics for Comparing Conformers. Two
metrics for comparison of conformer ensembles to an
experimental conformation were chosen, rmsd and the
Tanimoto combo, calculated using the Shape Toolkit from
OpenEye Scientific Software. The Tanimoto combo (or TC)
score represents a combination of shape matching (shape
Tanimoto) and functional group matching (color Tanimoto).
TC offers a complementary approach to the atom-based rmsd
measure for a number of reasons treated at greater length in
the Discussion section. An important difference between TC
and rmsd is that TC is bounded (by 0 and 2). It should be
noted that the processes used to overlay and score conformers
using rmsd or TC are very different, and therefore, the
conformer giving the minimum rmsd to the experimental
structure will almost always be different than the conformer
giving the maximum TC to the experimental structure.
Identifying Suitable Small Molecule Structures. Two
sources of small molecule structures, the PDB and the CSD,
were utilized. To find small molecules from the PDB, first,
suitable cocrystal structures were identified from larger sets
using global criteria; then, the ligands were extracted and
inappropriate ligands were removed using local, ligand-
centric criteria. Small molecule structures from the CSD were
obtained from a previously published data set,18which was
subjected to some of the ligand-centric criteria used to
identify suitable PDB ligands.
Identifying Suitable Cocomplexes from the PDB. There
are two kinds of criteria that must be considered here: global
criteria for the structure as a whole and local criteria
pertaining to the fit of the ligand to its local density and the
properties of the ligand itself. A number of possible criteria
of quality for crystal structures were enumerated in a paper
by Hartshorn et al. on the selection of a set of structures
suitable for validating docking programs for pose predic-
tion.20The global criteria identified by Hartshorn et al. were
that all structures must have good nominal resolution and
have deposited structure factors, while the local criteria were
that the ligand must be well fit to its local density and cannot
be covalently bound to the protein. This list has been
extended in this work, both at the global level and the local
level. In addition to the Hartshorn criteria at the global level
(good nominal resolution, deposited structure factors in the
PDB), the model must have good overall metrics of model
quality and reasonable coordinate precision. At the local
level, it is demanded that a ligand be well fit (using different
criteria to Hartshorn et al., vide infra) and be noncovalent.
It is further required that the ligands show no intra- or
intermolecular clashes and that they be diverse at the graph
level.
Nominal resolution, while very commonly used as a
measure of structure quality, is in fact a measure of the
quantity of data gathered, not quality of the structure, and
as such can only be used as a very rough guide to selecting
a good structure.21In many studies in this area, the nominal
resolution has been used as a surrogate for experimental
accuracy, which is inappropriate (see ref 22 and the Discus-
574
J. Chem. Inf. Model., Vol. 50, No. 4, 2010
HAWKINS ET AL.

Page 4

sion). However, since nominal resolution indicates the
quantity of data gathered, at a resolution of 2.7 Å or better,
the ratio of experimental data points to parameters in the
model is greater than 1. As such, structures with a resolution
of 2.7 Å or better offer the possibility (but only the
possibility) of a well determined model. The quality of a
deposited model can be independently checked only if
structure factors (electron density maps) are available in the
PDB, so all structures must have deposited structure factors.
In order to ensure that the models used are globally well fit
to the data (and not overfit), we use the difference between
the R-factor and the Rfree, as introduced by Brunger.23In this
work, we demand that a structure must have a Rfreewithin
0.05 of the R-factor.24In comparing computed and experi-
mental poses for ligands, it is almost uniformly assumed that
the experimental pose is infinitely precise and without error,
which is untrue for any piece of experimental data. The
average positional errors for atoms in a crystal structure can
be assessed using the diffraction-component precision index
(DPI).25We calculate DPI according to the approximation
published by Blow26(using the Blow rearrangement allows
us to calculate an estimate for DPI based solely upon
information commonly contained in the header of a PDB
file). The comparison of a computed conformer to the
experimental conformer by rmsd obviously requires com-
parison of atom positions between the two conformers.
Accordingly, the positional uncertainty in the atoms in the
experimental structures should be less than the rmsd for the
comparison to be meaningful. In this work, we estimate
positional uncertainty in atomic positions as sqrt(2) × DPI,
following Goto et al.27The maximum for the average
positional uncertainty in a structure was set at 0.6 Å, which
sets the maximum for the DPI to be 0.42 Å. In this way,
any rmsd result from this set greater than 0.6 Å is guaranteed
to be a prediction greater than the error in the experimental
data, and therefore, such a prediction will be significant.
Whatever the magnitude of the rmsd, a correction to it will
have to be made to account for positional uncertainty in the
ligand atoms (see the Results section). The assumption used
in this paper is that the ligand atoms have average positional
uncertainty. While the real uncertainty for the ligand atoms
may well be different, it is not reliably calculable at the
present time; therefore, we use DPI as our best estimate.
When considering whether or not a ligand is suitable for
conformer generation studies, the local quality of fit of the
ligand to its density, as well as global metrics of quality,
should be considered. In this study, three local metrics, the
real-space correlation coefficient (RSCC),28the real-space
R-value (RSR),29and the occupancy-weighted B-factor
(OWAB) were used to ensure that the ligand has been
sensibly fitted to the electron density. The use of the
occupancy weighted B-factor as a criterion was based not
on the idea that this number necessarily measures thermal
mobility in the ligand, as is often assumed. Rather, we used
the OWAB to indicate either structures where the ligand is
genuinely disordered in the active site, and therefore, the
idea of reproducing “the” bioactive conformation is question-
able or structures where there is some pathology of the model
that led, in the process of fitting the ligand to its density, to
very high or very low B-factors. These models are inherently
of dubious value and should not be included in the set. The
criteria used in this work were that RSCC > 0.9, RSR < 0.2,
and 1 < OWAB < 50. These data were obtained by download
from the electron density server (EDS).30
When applied sequentially, these global and local filters
should ensure that a ligand structure arises from a good
quality overall model and has been well fit to its local density.
The filters were applied to a large set of PDB structures (see
Results), and the cocomplexes that survived were then
processed to separate the ligand and the protein. These pairs
were checked to ensure that they do not show any intermo-
lecular clashes, and those ligands that showed no clashes
were designated as well-solved structures, to be used in the
final, ligand-based filters. In these final steps, we focus on
identifying suitable molecules using simple physicochemical
and graph-based properties.
Identifying Suitable Ligands from the PDB. In the last
filtering step, we removed ligands that are not representative
of the population of molecules for which this validation is
intended. As a first step, we removed all well-known
cofactors, as these were not as relevant to the goals of this
study (validation of OMEGA’s performance on druglike
molecules). As mentioned above, OMEGA depends upon
MMFF94 for its energy calculations, so molecules that
contain elements unknown to MMFF94 (such as boron) must
be removed. To ensure that we used a collection of ligands
that provides as broad a set of tests for the OMEGA
algorithm as possible, we attempted to select a structurally
diverse set of ligands. At the simplest level, the molecules
can be neither too rigid (so that the conformer sampling
problem is trivial) nor too flexible (so that a close reproduc-
tion of the experimental conformation is more a matter of
good luck than a good algorithm because of the high
dimensional nature of the search space) and neither too small
nor too large. We, therefore, applied the following criteria:
3 e maximum number of rotatable bonds e 16; 8 e heavy
atom count e 50.
Preliminary analysis revealed that molecules designed to
bind to thrombin were over-represented in the set of well-
solved ligands identified by the criteria used thus far. These
ligands tended to share a high degree of substructural
similarity, so a method to introduce some level of structural
diversity at the graph level was sought. We applied a simple
measure of molecular similarity using the LINGOS method;31
any pair of molecules that had a LINGOS Tanimoto >0.9
was flagged for removal, and the molecule with the better
set of local fitting metrics for the pair (RSCC, RSR and
OWAB) was retained. The use of graph-based metrics like
LINGOS for structural diversity is a surrogate for the more
relevant but more difficult to assess diversity in the torsions
and flexible rings present in the molecule.
To further reduce the influence of experimental error on
this data set, we demand that the ligand does not show any
nonbonded intramolecular atomic clashes, as these are a mark
of a poorly solved model. As further discussed in the Results
section, we turned to the CSD for precise distributions of
interatomic distances with which to identify clashing atoms.
A final visual check was performed to remove molecules
that are structurally analogous but were not identified as such
by the LINGOS comparison (those with large common
substructures and/or a large fraction of their functional groups
in common). This gave us 197 ligands whose conformations
we then attempted to reproduce using OMEGA. The attrition
rates for the different data sets and an analysis of OMEGA’s
CONFORMER GENERATION WITH OMEGA
J. Chem. Inf. Model., Vol. 50, No. 4, 2010 575

Page 5

ability to reproduce these 197 surviving structures is given
in the Results section.
A parallel analysis was performed on high quality small
molecule structures from the CSD, as outlined in the
following paragraphs.
Identifying Suitable Molecules from the CSD. An
important part of the selection criteria for structures from
the PDB focused on removing structures that were not
reliably and accurately fit to their density, and a large
majority of the candidate structures failed these criteria (see
the Results section). A further problem with molecules from
the PDB is the high level of inherent experimental uncertainty
in the atomic positions. Both of these problems can be
avoided entirely if small molecule structures from the
Cambridge Structural Database (CSD) are used. A high
quality set of 492 druglike structures has been extracted from
the CSD as part of a previously published study,18and these
were used as the basis for generating a complementary set
of structures to those we obtained from the PDB. No filtering
of these structures based on model fit (global or local) was
necessary as all atoms are accurately located in these models.
Applying the same physicochemical and diversity criteria
as were used for the ligands from the PDB gave a set of
480 small molecule structures upon which we could perform
the same validation experiments. Comparison of these results
to those from the ligands from the PDB can be found in the
Results section.
All code was written in Python (version 2.4), and statistical
calculations were performed with Rpy version 1.03. Chem-
informatics functions (including rmsd and LINGOS calcula-
tions) were performed using OpenEye’s OEChem (version
1.6.1), and shape calculations were performed using the
OpenEye Shape toolkit (version 1.6). Protein-ligand coc-
rystal structures were downloaded from the PDB and filtered
according to the criteria laid out in the Methods section.
RSCC, RSR, and OWAB data for each PDB ligand were
obtained from the EDS. Molecules from the CSD were
obtained from one of the authors of the original study
(personal communication, K. Brameld).
Molecules were converted into isomeric SMILES format
using the OEChem toolkit before being used for conformer
generation. Conformer databases were generated using
OMEGA version 2.3, with an energy window for acceptable
conformers of 10 kcal/mol above the ground state using a
modified version of the MMFF94 force field, a maximum
number of conformations per molecule of 200, and an rmsd
cutoff of 0.5 Å (the default settings in OMEGA2.3).
Interatomic distance analysis of the CSD was carried out
with ConQuest v.1.6.32
RESULTS
Molecules from the PDB. In the Methods section, we
presented a set of criteria for selecting suitable protein-ligand
cocomplexes from the PDB and then identifying suitable
ligands from those complexes. We applied these criteria to
three large data sets (two of which have previously been
used in conformer validation studies) that together consist
of over 4500 cocrystal structures: 778 structures that Kirch-
mair et al. used in their comparison of Catalyst and
OMEGA,17the set used by Sadowski and Bostrom in their
study on the OMEGA 1.8.1 torsion library13(1267 structures)
and a subset of the PDBBind database16(2516 structures).
The percentages of structures remaining after each filter was
applied are illustrated in Figure 1. In all three databases, we
found very large levels of loss when we applied our filtering
criteria (over 90% of the structures in every database were
removed). It can easily be seen from Figure 1 that there are
three criteria that, when applied in the given order, remove
the highest percentages of the structures: nominal resolution
greater than 2.7 Å (criterion A), poor fits of the ligand to its
density (criterion E), and inappropriate physicochemical
properties (criterion F).
The problem of identifying atom-atom clashes (filters E
and G) based on analysis of interatomic distances in
structures from the PDB is nontrivial, due to the inherent
uncertainty in those atom positions. Therefore, we chose to
perform an analysis of nonbonded contacts for commonly
occurring atoms (C, N, O, S, P, F, Cl) using high quality
structures (R-factor <0.05, no disorder) of organic molecules
from the CSD, where the atoms are located with very high
precision. The cutoffs identifying clashes were set at an
interatomic distance above which 95% of the distances for
that (nonbonded) atom pair in the CSD occurred. These
cutoffs are shown in Table 1 in the Supporting Information.
The only well fit ligands that can show intermolecular
clashes are those that are covalently bound to their target
protein, and these are explicitly excluded from this set.
Therefore, noncovalent ligands that have clashes, by these
criteria, with their cognate protein are incorrectly solved.
Figure 1. Attrition rates for three data sets of cocrystallized ligands. The filters used are (A) nominal resolution e2.7 Å; (B) structure
factors must be present; (C) R - Rfreemust be <0.05; (D) DPI < 0.42 Å; (E) RSCC > 0.9, RSR < 0.2, OWAB < 50, no intermolecular
clashes; (F) physicochemical properties; (G) pairwise LINGOS similarity <0.9, no intramolecular clashes.
576
J. Chem. Inf. Model., Vol. 50, No. 4, 2010
HAWKINS ET AL.