Comparison of Shape-Matching and Docking as Virtual Screening Tools
Paul C. D. Hawkins,* A. Geoffrey Skillman, and Anthony Nicholls
OpenEye Scientific Software, 3600 Cerrillos Road, Suite 1107, Santa Fe, New Mexico 87507
ReceiVed March 22, 2006
Ligand docking is a widely used approach in virtual screening. In recent years a large number of publications
have appeared in which docking tools are compared and evaluated for their effectiveness in virtual screening
against a wide variety of protein targets. These studies have shown that the effectiveness of docking in
virtual screening is highly variable due to a large number of possible confounding factors. Another class of
method that has shown promise in virtual screening is the shape-based, ligand-centric approach. Several
direct comparisons of docking with the shape-based tool ROCS have been conducted using data sets from
some of these recent docking publications. The results show that a shape-based, ligand-centric approach is
more consistent than, and often superior to, the protein-centric approach taken by docking.
In recent years virtual screening (VS) has become an
important part of the armamentarium of modern drug discovery.
Much of the drive to use virtual screening has arisen from
increased pressure to put compounds into the development
pipeline and to reduce the costs of getting suitable compounds
to this point. Given the well-known costs involved in experi-
mental high-throughput screening (HTS),1virtual screening has
been increasingly applied to reduce the number of compounds
going into experimental HTS. For the purposes of this discussion
we define virtual screening as ranking molecules by descending
order of likelihood of relevant biological activity, regardless of
how that ranking is performed.2Given that the time and costs
associated with HTS can be reduced by correctly applied virtual
screening, much effort has been expended in identifying VS
approaches that assign low ranks to most of the inactive
compounds while assigning high ranks to most or all of the
Structure-based virtual screening approaches fall into two
main classes: those based on protein coordinates and those
based on ligand coordinates. When a protein-ligand cocrystal
is available in an industrial project setting, virtual screening
using docking has become the method of choice.3This may
reflect the preponderance of recent published work on virtual
screening via docking and scoring. In addition, there are now
several “distributed processing” initiatives, i.e., virtual screening
projects using spare cycles on personal computers, based on
docking (FightAIDS@Home, ScreenSaver/LifeSaver, D2OL).
Some notable exceptions to this trend have focused on phar-
macophore or QSAR-based approaches, often, though not
exclusively, in systems where structural information on the
protein-ligand complex is not available.4-7For recent reviews
on structure-based virtual screening, see the articles by Lyne8
Docking can be divided into two parts: the correct positioning
of the correct conformer of a ligand in the context of a binding
site (posing) and its successful recognition/scoring by a scoring
function (scoring). Such an approach is an attempt to simulta-
neously solve several difficult problems and often results in
inconsistent performance.10The reasons for this are manifold
and include (i) inadequate treatment of electrostatics, electronic
polarization, aqueous desolvation, and ionic influences, (ii) lack
of accounting for entropy changes in the protein and the ligand
on binding, (iii) insufficient ligand conformer sampling, (iv)
inadequate sampling and weighting of proton positions (tau-
tomers, rotamers) and charge states (ionization) of both protein
and ligand, (v) the assumption of a rigid protein, and (vi) the
well-known deficiencies in the scoring functions used to rank
the docked molecules.10,11
A few publications on structure-based virtual screening have
been prospective12and have clearly demonstrated the strengths
and the limitations of the technique. The work of Jenkins12a
illustrates the importance of consensus methods in ranking
compounds and the weaknesses of a single docking/scoring
combination, as the performance of any one docking engine/
scoring function was never found to be consistently optimal.
The work of Forino,12bwhile providing disappointingly low hit
rates, clearly shows that docking can be of use in identifying
compounds that are selective for the target of interest over
related proteins. In some cases prospective docking studies have
been performed that have provided a direct comparison with
experimental approaches,13,14while others have used docking
in concert with ligand-based approaches.1,15
Most of the structure-based virtual screening papers, however,
are retrospective evaluations or comparisons of two or more
docking tools, using a small number of known binders to the
target of interest (actives) placed into a database of compounds
that are presumed to be inactive (decoys). The performance of
the tools under examination is then quantitated by some metric
related to the ranking of the actives versus the decoys. Since a
number of these publications provide the data sets used in the
evaluation as Supporting Information, the potential arises to
compare directly the performance of these structure-based
approaches to techniques based purely on the ligand. Accord-
ingly we present data directly comparing the performance of
structure-based methods and a shape-based, ligand-centric
method (ROCS) on the same data sets, allowing the most direct
and meaningful comparisons to be made. In this work we rank
molecules on the basis of their similarity to a known active
molecule in three-dimensional shape space, using atom-centered
Gaussian functions to allow rapid maximization of molecular
overlap.16A more extensive presentation on shape as a metric
property for molecular comparison is contained in Haigh et al.17
For a comprehensive listing of tools to align or overlay
molecules based on a wide variety of metrics, see the work of
* To whom correspondence should be addressed. Phone: 505-473-7385,
extension 65. Fax: 505-473-0833. E-mail: firstname.lastname@example.org.
J. Med. Chem. 2007, 50, 74-82
10.1021/jm0603365 CCC: $37.00 © 2007 American Chemical Society
Published on Web 12/13/2006
Melani et al.18Lemmen and Lengauer19have extensively
reviewed the use of ligand-based tools in virtual screening.
The comparisons in this paper focus on data provided in
publications from Sanofi-Aventis,20Johnson & Johnson,21the
Jain lab,22the Villoutreix lab,23and the Rognan lab.24In the
work from Sanofi-Aventis docking was performed into homo-
logy models, and in the other publications docking was
performed into experimental crystal structures. The studies from
Sanofi-Aventis, Johnson & Johnson, the Villoutreix lab, and
the Rognan lab have been replicated in their entirety with a
shape-based approach, whereas in the case of the Jain data, only
those cases in which at least 10 actives were provided for a
given target are replicated. This is to ensure that the results have
some statistical meaning. We note that there are operational
difficulties in comparing applications for virtual screening in a
statistically meaningful way, especially when there are only a
handful of active compounds in the database being searched.25
The data presented herein were obtained using the same
starting point as that for docking: a protein-ligand cocrystal.
In docking, the ligand is extracted from the complex and
discarded and then attention is focused entirely on the volume
in the protein active site revealed by removal of the ligand. For
shape-based similarity, however, it is the protein that is discarded
and attention focused on identifying compounds that best match
the volume and disposition of functional groups in the ligand.
In this study the query ligand was used in its experimental
conformation whenever that is available.
Data sets were obtained as Supporting Information to the
publications21,22or directly from the authors.20,23,24Databases
of conformers for the data sets were then created using
OMEGA.26The ligand structures used as queries were extracted
from experimental cocrystal structures (obtained from the
PDB27) and processed using OEChem.28The ligand was then
used in its crystallographic conformation as the query for
ROCS.29In the work of Evers on G-protein-coupled receptors
(GPCRs),20where no crystallographic conformations for the
ligands are available, a single low-energy conformation for the
query molecule was generated using OMEGA.
ROCS performs shape-based overlays of conformers of a
candidate molecule to a query molecule in one or more
conformations. The overlays can be performed very quickly
based on a description of the molecules as atom-centered
Gaussian functions. ROCS maximizes the rigid overlap of these
Gaussian functions and thereby maximizes the shared volume
between a query molecule and a single conformation of a
In preliminary work the effects of various options available
in ROCS were examined for their impact on the VS performance
of ROCS. In default operation ROCS compares molecules based
purely on their best shape overlap, quantitated by their shape
Tanimoto. ROCS then ranks the database molecules based on
their shape Tanimoto to the query molecule. It was quickly
found that adding to the shape Tanimoto the score for the
appropriate overlap of groups with like properties (donor,
acceptor, hydrophobe, cation, anion, and ring), the so-called
color score, and then ranking on this summed score improved
virtual screening performance considerably. Donors and accep-
tors were defined according to Mills and Dean.30Cations and
anions were defined according to an implicit pKamodel such
that the same group (e.g., carboxyl group) had the same
protonation state (e.g., ionized) regardless of the protonation
state set in the input structure definition.
Given the marked improvement when using the color score,
all our experiments were performed with ROCS in “color-
optimization” mode. In this mode ROCS optimizes the molec-
ular overlay to maximize both the shape overlap and the color
overlap obtained by aligning groups with the same properties
that are contained in the color force field file. This overlay is
then subsequently scored using the sum of shape Tanimoto for
the overlay and the color score (the so-called combo score).
Customization or target-specific information can be incorporated
by adding a term to the color force field file that rewards overlay
of specific functional groups. For example, groups that might
be required for tight binding to the protein in question might
be given extra weight in the color file (e.g., amidines/guanidines
for thrombin). For an application of this approach, see the results
from the comparison with the work of Evers et al.20below.
There are a number of approaches to quantitating the success
of a particular tool for virtual screening. The most often used,
and simplest to calculate, is enrichment at a given percentage
of the database screened. Enrichment (EF) is defined as
where Hitssampledx%is the number of hits found at x% of the
database screened, Nsampledx%is the number of compounds
screened at x% of the database, Hitstotalis the number of actives
in entire database, and Ntotal is the number of compounds in
entire database. It can easily be seen that enrichment has a fixed
maximum at any given percentage of the database screened.
At 1%, the maximum is 100, at 2% the maximum is 50, and at
10% screened the maximum enrichment obtainable is 10.
Another technique that portrays success at fixed percentages
through the database is the hit rate (HR), the percentage of
the known hits that are contained in a set percentage of the
where Hitssampledx%and Hitstotalare defined as above.
These metrics both rely on cutoffs made at various points
through the ranking and so can be sensitive to small changes in
ranking. A variant of the enrichment statistic has been developed
to avoid this sensitivity, the so-called robust initial enhancement
(RIE) approach.31A measure that assesses virtual screening
performance across the entire database and hence is not sensitive
to small changes in ranking is the area under the receiver
operator characteristic, or ROC. The ROC shows performance
of a given tool when screening across the entire database is
examined and not just at fixed, early points in the screen as
enrichment, hit rate, and to some extent RIE do. The theoreti-
cally perfect performance of a virtual screening application gives
the maximum area under a ROC curve (1.0), while random
performance of a tool gives an area under the curve (AUC) of
0.5. AUC values of less than 0.5 imply a systematic ranking of
decoys higher than the rankings of known actives. The ROC
approach has been used extensively in the life sciences and
social sciences arenas since its inception.32,33For a recent
application of the ROC curve in virtual screening, see the work
In the following sections the performance of ROCS will be
compared to the performance of a variety of docking tools. In
Shape-Matching and Docking Journal of Medicinal Chemistry, 2007, Vol. 50, No. 1 75
each experiment a shape and chemistry similarity based screen
was performed with ROCS, based on similarity to a query
molecule. In most cases this query molecule is the crystal-
lographic ligand from the same cocomplex that was used in
the docking experiments. In the case of the work on virtual
screening against GPCRs, no crystal structure information is
available, so the query molecule was selected in a different
manner (vide infra). Tables showing the fingerprint similarities
(Tanimoto coefficient based on MACCS keys) of the known
active molecules being searched for to the query molecule can
be found in the Supporting Information.
The Work of Bissantz. Bissantz et al. examined three well-
known docking tools (FlexX, GOLD, and DOCK) as virtual
screening tools against thymidine kinase (PDB code 1kim) and
the estrogen receptor (PDB code 3ert).24The database being
screened consisted of 990 decoy molecules chosen from the
ACD, seeded with 10 TK actives and 10 ER actives. The study
was initiated by identifying the appropriate scoring function(s)
that most effectively discriminated the known actives from the
decoys. With this knowledge in hand, the authors then performed
the retrospective virtual screen and presented the results in terms
of the enrichment and hit rate (vide supra) at 5% of the database
screened. A comparison of enrichments between the docking
tools and ROCS is illustrated in Figure 1. Note that the maximal
enrichment that can be achieved at 5% of the database screened
is 20. In both cases the hit rates from docking are obtained by
a consensus score using two scoring functions that had been
identified as effective against the target in question.
It should be noted that the performance of the docking tools
is somewhat variable across the two targets. It has frequently
been observed that the binding site of the estrogen receptor is
highly hydrophobic, rigid, and sterically constrained, making
it an easy target for docking. In contrast, the thymidine kinase
binding site is hydrophilic with water molecules making bridging
interactions between the ligand and the protein. There is also a
flexible loop making up one side of the site, making it a
challenging target for docking. Some or all of these differences
between the sites may explain the reduced performance of
GOLD and FlexX against TK (with GOLD showing the most
significant decrease). ROCS, in contrast, shows consistently
good performance on both targets when using the same ranking
method in each case (the combo score).
The Work of Evers. Given the high importance of GPCRs
as drug targets, where 9 of the top 20 best-selling prescription
drugs in 2000 targeted GPCRs,35much attention has been
focused on effective virtual screening methods for this target
class. Docking has not seemed to be a promising tool for virtual
screening against GPCRs, given the lack of any high-resolution
structure of any human GPCR.
In an attempt to determine the effectiveness of building
homology models of GPCRs and then performing docking,
Evers et al. performed virtual screening by docking into a
homology model of the R1-adrenoreceptor.36An extensive study
was conducted to determine the scoring function that was most
effective in discriminating actives from decoys after docking
with GOLD. It was shown that a high proportion (5 of 9) of
the scoring functions examined were unable to perform better
than random selection at up to 10% of the database screened,
while 2 of 9 were only little better than random at the same
point. Both this study and the work of Bissantz24illustrate that
a considerable amount of effort can be invested in identifying
the appropriate scoring function(s) for a given combination of
docking engine and target.
To test the performance of ROCS in GPCR virtual screening,
we turned to related work from Evers that compared virtual
screening performance against four biogenic amine binding
GPCRs (R1A, 5HT2A, D2 and M1).20In this work several tools
were utilized, including docking into homology models (with
GOLD and FlexX-Pharm), pharmacophore tools (Catalyst), 3D
similarity searching (FlexS), and a variety of 2D approaches.
The virtual screening was conducted on a database consisting
of 50 diverse active compounds and 950 diverse decoys, all
selected from the MDDR.37Note that in the course of this work
GOLD and FlexX-Pharm were used with constraints based both
on the positions of protein atoms (side chains that are known
to make contact with active ligands) and of ligand atoms (the
protonatable nitrogen that is known to be essential for activity
against these targets). To make the fairest comparison, we
elected to try to mimic these types of constraints in ROCS.
Lacking the homology models used in the paper, we could not
utilize the protein-based positioning constraint. Rather, an
additional term was added to the color force field file that
rewarded placing protonatable nitrogens atop one another in the
overlays, in an attempt to mimic the effect of the constraint in
docking. Note, however, that the constraint in docking is an
absolute constraint; failing to match it will result in a pose being
rejected. In contrast, in ROCS an overlay not providing
appropriate alignment of the protonatable nitrogens will not be
rejected; it will simply not accrue the extra score. In the ROCS
experiments the query molecule was the same molecule that
was used in the FlexS portion of the study, and as mentioned
above, a single conformation for these molecules was generated
with OMEGA. Figure 2 shows comparisons of the performance
of the 3D techniques in the paper with ROCS.
It can be seen that in two cases, 5HT2A and D2, ROCS with
the amended color force field file performs very well compared
to the other tools. For A1A, ROCS performs a little better than
the other tools, while against M1, Catalyst or ROCS would be
For virtual screening against the A1A receptor, Evers
provided data on the performance of GOLD with and without
the protein-based and ligand-based constraints mentioned
above. Unsurprisingly, GOLD’s performance is significantly
improved by the addition of these constraints (see Figure 3).
The use of a color force field file in ROCS that rewards over-
lap of protonatable nitrogens (thereby mimicking one of the
Figure 1. Hit rates for docking tools and ROCS against two targets,
the estrogen receptor (ER) and thymidine kinase (TK). PDB codes for
the docking targets are 3ERT (ER) and 1KIM (TK).
Journal of Medicinal Chemistry, 2007, Vol. 50, No. 1Hawkins et al.
constraints used in GOLD and FlexX-Pharm) provides an
improvement over the default file only at 10% of the database
In a comparison of the data from Catalyst, it is worth noting
that multiple active compounds (up to 20) were used to develop
the pharmacophores used for searching rather than the single
molecule used in the ROCS and FlexS portions of the study.
On the basis of the much superior performance of Catalyst to
FlexS, Evers et al. state, “Accordingly, the good performance
[of 3D ligand-based methods] must be attributed to the fact that
a wide range of structurally diverse active compounds for each
target is available.” However, given that overall ROCS outper-
forms Catalyst quite significantly in these examples, this
conclusion must be questioned. ROCS has performed well when
using only a single, low-energy conformation of a single
molecule as a query.
The Work of Cummings. Cummings et al. examined four
docking tools (DOCK, DOCKVISION, GLIDE, and GOLD)
for use against three publicly available targets (HIV-1 protease,
protein tyrosine phosphatase-1B, and thrombin).21In each case
there are 10 active compounds placed into a background of 990
decoy compounds from the MDDR.37
The paper utilizes enrichment at three points through the
database (2%, 5%, and 10%) as its performance metric. Figure
4 shows the comparison of ROCS to the docking tools
investigated in the paper. Note that while some docking tools
perform very well on certain targets (e.g., GOLD on PTB-1B,
especially at the 2% point), there is no tool that is more
consistent than ROCS, and only GLIDE shows performance of
equivalent consistency. In the case of HIV-PR DOCK, DOCK-
VISION and GOLD were unable to identify a single active even
in the top ranked 10% of the database, accordingly giving no
enrichment at up to 10% of the database screened. The extreme
difficulty that most of the docking tools have with HIV-1
Figure 2. Performance of ROCS and other tools in virtual screening
against four GPCRs. ROCS_cff denotes ROCS performance with the
amended color force field file.
Figure 3. Comparison of GOLD and ROCS performance with and
without constraints. GOLD_pc shows performance with constraints.
GOLD shows GOLD default performance. ROCS_cff shows ROCS
performance with an amended color file. ROCS shows ROCS default
Figure 4. Comparison of enrichment at 2%, 5%, and 10% for docking
tools and ROCS. The PDB codes for the crystal structures used in the
docking are 1HVR (HIV-PR), 1C84 (PTP-1B), and 1QBV (thrombin).
Shape-Matching and Docking Journal of Medicinal Chemistry, 2007, Vol. 50, No. 1 77
protease most probably arises from the large size and high
flexibility of many HIV-1 protease ligands.
In the cases illustrated so far the experiments have been on
a relatively small scale (1000 molecules in total, between 10
and 50 actives). Now we discuss a larger scale experiment that
was recently disclosed.
The Work of Miteva. In a large-scale virtual screening
experiment Miteva et al. used a two-step procedure for docking,
first using FRED38as a fast, shape-based filter to remove
compounds that cannot fit into the target protein’s binding site
and then docking and ranking the remaining compounds with
DOCK or Surflex. They examined four targets: the estrogen
receptor, thymidine kinase, neuraminidase, and factor VII. The
decoy compounds were selected by removing unsuitable com-
pounds from the ACD, leaving 65 611 “druglike” compounds.
The actives were taken from the PDB (10 each for ER, TK,
and NA and 19 for factor VII), to give a total database size of
A comparison of these docking approaches to ROCS is shown
in Figure 5. The data are presented as a plot of percentage of
the database screened versus percentage of the known actives
identified. It should be noted that the sizes of the databases
screened are different, as the database used for the ROCS study
was not prefiltered by FRED, so that the entire database of over
65 000 compounds was screened. In the case of DOCK and
Surflex the databases are considerably smaller (between 15 000
and 30 000 compounds depending on the target) due to the
prefiltering performed by FRED.
In a note on timing in their paper, Miteva et al. observe that
the average time per ligand when docking with Surflex is around
10 s, and with DOCK around 8 s, on a 1.5 GB RAM, 2.8 GHz
Xeon processor. In this study performing ROCS overlays on
the 65 660 compounds required an average over the four targets
of 0.25 s per ligand. The average time to make conformers for
this database was 2.1 s per molecule. These timings are for a 1
GB RAM, 1.0 GHz Pentium 3 processor. Clearly this protocol
is significantly faster than the docking part of the approach
documented by Miteva.
Note that in all cases ROCS identifies all the actives in only
8% of the database whereas the docking tools are unable to do
this in all cases. Surflex fails to identify all the actives for TK
and F7, while DOCK fails to identify all actives for TK, ER,
and NA. Possibly some of the actives are too large to fit into
the rigid active site representation used by these docking tools,
and hence, they fail to give scores. DOCK can only identify all
the actives in one case, factor VII, but in this case it does so
with remarkable efficiency, ranking all the actives in the top
0.6% of the database. Overall in this comparison ROCS shows
itself to be a more efficient tool than docking by identifying,
on average, a larger proportion of the actives at a fixed
proportion of the database.
Further examination of the ranked lists from ROCS allows
us to determine what fraction of the database must be screened
before at least one example of every chemotype of inhibitor is
found. For neuraminidase this fraction is 0.014%, and for the
estrogen receptor the fraction is 0.36%. The factor VII inhibitors
lie essentially in one structural class, and the answer is obscured
in the case of TK by compounds very similar or identical to
the actives that are present in the decoy set. Some of the pitfalls
in this study that were encountered due to the nature of the TK
decoy compounds are elaborated in the Discussion.
Inspection of the 2D similarities of the query and active
molecules (see Table 3 in the Supporting Information) shows
that the success of the shape-based approach is not sensitive to
the 2D similarities of the query and the active compounds. For
example, ROCS performs very well against the estrogen receptor
and neuraminidase, and yet the actives used show only moderate
2D similarity to the query. A more thorough investigation of
the comparison between shape-based approaches and 2D
approaches in QSAR and ranking is found in ref 39.
The Work of Jain. In the work of Pham and Jain22the
performance of Surflex is illustrated using the area under the
receiver operator characteristic curve (vide supra). In the original
study 26 protein targets were investigated, while in this work,
a subset of targets for which at least 10 active molecules were
available was used (8 targets). Figure 6 shows a plot of the
area under the ROC curve for these eight targets from Surflex-
Dock and from ROCS
Figure 6 shows that the performance of Surflex-Dock and
ROCS is usually quite similar. In the case of trypsin, Surflex-
Dock is clearly the superior tool, while for thrombin and PARP,
ROCS is a superior choice. In all other cases the difference is
not significant. Possible reasons for the poor performance of
ROCS against trypsin are outlined in the Discussion. The
Figure 5. Plots for four targets comparing recovery of known actives
with percentage of database screened. Targets and PDB codes are as
follows; TK, thymidine kinase (1F4G); ER. estrogen receptor (3ERT);
F7, factor VII (1DVA); NA, neuraminidase (1B9S).
Journal of Medicinal Chemistry, 2007, Vol. 50, No. 1 Hawkins et al.
approach to docking taken by Surflex (matching a ligand to a
“proto-molecule” defined by the volume and features of the
active site) is somewhat similar in conception to ROCS. It is
therefore noteworthy that its results are so similar to those of
Conformer databases were generated with OMEGA 1.8.1,26using
an energy window for acceptable conformers of 8 kcal/mol above
the ground state and a rmsd cutoff of 0.8 Å (see Bostrom40for a
discussion of the effect of parameters on OMEGA performance).
The maximum number of rotatable bonds allowed in any molecule
was 16 (except for the case of the OPPA data set from the Pham
study where the maximum was set to 25), resulting in the loss of
a very small fraction of compounds from some of the data sets.
The conformer databases were searched with ROCS 22.214.171.124The
query molecule for each ROCS run was in a single conformation,
the conformation found in the protein-ligand X-ray structure in
the PDB.27In the cases where no cocrystal structure was available,
a single low-energy conformation was used as the query, generated
with OMEGA (by setting the maxconfs parameter to 1). ROCS
was run using a built-in color force field file (with the -chemff
ImplicitMillsDean flag), and all overlays were optimized to
maximize color overlap after the best shape overlay was located
(using the -optchem flag). The hits were ranked on the basis of
the sum of their shape Tanimoto and the normalized color score in
this optimized overlay (using the -rankby combo flag). This sum
is known as the combo score. The color force field file used in the
GPCR studies was amended to add the following terms to reward
overlap of protonatable nitrogens in the query molecule and
Note that the standard weight of color interactions is 1.0, so the
weight awarded to the overlap of these protonatable nitrogens is
considerable. This is done to mimic the use of pharmacophoric
protonatable nitrogens, where a molecule that cannot place the
appropriate atom in the correct place is rejected by the docking
Structure-based virtual screening is a well-known and widely
applied technique in modern lead discovery, most commonly
involving the procedure known as docking. Docking programs
are typically used to predict one of three things, in order of
decreasing difficulty: affinity, binding mode, and activity. There
is general agreement41that docking programs cannot predict
affinity to a degree that is useful. On the other hand, they have
shown some utility in predicting binding modes and i activity,
i.e., selection of active compounds from a larger set of inactives.
In this study we have focused on the relative merits of docking
and a shape-based, ligand-orientated method, ROCS, for the
last of these functions, separating the binders from nonbinders.
This study does not address the other two areas.
One criticism of such a comparison concerns explicit and
implicit parametrization. Docking methods typically contain
implicit parametrization via scoring functions. Because scoring
functions are developed on systems typically not dissimilar to
docking targets, there can be implicit parametrization toward
the known. Many in the field do not see this as a problem; in
fact, such “knowledge-based” potentials are very popular.
Ligand methods are typically explicitly parametrized; i.e., a
series of actives for a system are used to generate a query
intended to locate and rank highly other actives. In this respect
ligand-based methods fall midway between the pure 3D nature
of docking and the connection table methods common in most
QSAR. One of the advantages of the approach taken in this
study is that the shape-based approach does not require multiple
active compounds and, as such, requires no explicit parametriza-
tion. This work includes one exception, to favor protonatable
nitrogens as ligands for amine binding GPCRs (vide infra) but
only because in this case docking also included such an explicit
In order the minimize the possibility of local knowledge bias
common in the field, we have extracted five data sets from the
literature and replicated the docking experiment, using the
combination of shape and chemical similarity as the ranking
method. The results lead us to conclude that the shape-based
approach can provide better performance than docking tools in
more than half of the 21 systems examined. We also note that
knowledge of the bioactive conformation of the query molecule
is not necessary for the shape-based approach to give good
The outcome of the experiments detailed above was unex-
pected. Shape similarity is not a profound technology, it merely
aligns volumes and adds in a term for functional group
similarity, and equal weight in the final score is given to both
the shape and chemical similarity contributions. It is likely that
extensive parametrization would show that assignment of
different weights to the shape and color parts of the combo score
could provide superior performance in virtual screening or other
applications. However, this has not been done, and given the
successes illustrated in this paper, it seems that the naive
approach of assigning exactly equal weights to the shape and
chemical similarity components of the combo score was
justified. That such a straightforward approach should provide
equivalent performance at the simplest task asked of docking
(ranking) to more sophisticated methods with many years of
investigation behind them requires comment. To fulfill its
promise, docking needs to accurately predict protein-ligand
interactions, something not yet possible. In place of these
predictions are heuristic scoring functions that attempt to capture
some of the essence of binding physics. That docking works at
all is a triumph of such functions. However, scoring functions
are also notorious in promoting false positives. It is not that
such functions do not recognize active ligands and predict
binding modes, but they cannot recognize inactive molecules.
The necessity of including information on bad ligands, as well
as good, in scoring function development is a major thrust of
the work of the Pham and Jain.22
Figure 6. Comparison of the area under the ROC curve for eight targets
using Surflex-Dock or ROCS: ROCS, result when using the crystal-
lographic pose of the query. The PDB codes for the proteins are as
follows: OPPA, 1B5J; trypsin, 1QBO; HIV protease, 1PRO; TK,
1KIM; PTP-1B, 1PTY; PARP, 2PAX; thrombin, 1C4V; TS, 1F4G.
DEFINE CATaventis [N;!$(*CdO);!$(a*(c)c);!$(*[+])]
PATTERN gpcr [$CATaventis]
INTERACTION gpcr gpcr attractive gaussian weight)
Shape-Matching and Docking Journal of Medicinal Chemistry, 2007, Vol. 50, No. 1 79
Shape-based approaches often suffer from the inverse issue,
the problem of the false negative. The underlying assumption
of shape-based methods is that compounds with shape and
chemistry similar to those of a known active molecule have a
high probability of also being active. Consequently active
molecules with shapes different from that of the active used as
a query could easily be missed. Accordingly, one could imagine
docking finding radically different ligands, in size or shape that
shape similarity is unable to capture, allowing the identification
of novel ligands with unforeseen binding interactions. However
the rigid protein assumption employed by all docking engines
in this study (and other approximations mentioned above) often
means that this promise is not fulfilled. We see an example of
the size bias in shape-based approaches when examining trypsin
in the Jain experiment, where Surflex-Dock gave much better
performance than ROCS. Here, the ligand from the PDB
structure 1QBO, used as the query, is relatively large, whereas
a high proportion of the active ligands in this data set are quite
small. The ROCS algorithm begins the shape overlay procedure
by overlapping the centers of mass of the query and the database
conformer and then aligning their principal moments of inertia.
A consequence of this is that if two molecules of very different
sizes, but possessing some of the same functional group(s), are
aligned on the basis of the moments of inertia, then the
functional groups will quite likely not be aligned at all, and the
overlay solution may be trapped in a local minimum of the shape
and color force hypersurface. Therefore, the combined shape
and chemistry (color) score for the molecules will be low.
However, such failures do not dominate the overall statistics
and can be addressed by including alternative queries or using
asymmetric measures of shape similarity rather than the sym-
metric Tanimoto measure used here.
This observation does lead to one of the confounding
questions of ligand-based design: Which compound(s) (and in
which conformations) should be used as the query, and how
should they be chosen? In our studies the ROCS approach to
shape similarity has proven to be extremely robust to ligand
choice. Only one molecule was used as the query (unlike the
approach routinely used in pharmacophore tools where multiple
active molecules are required). The shape-based approach
routinely crosses boundaries in chemical space, as shown in a
follow-up to a recent study by Warren et al.10A comprehensive
comparison was made on 10 commercial docking programs
using almost 1300 ligands from 21 chemical classes against 8
protein targets. Each target had crystal structures of ligands from
multiple chemical classes. The authors kindly compared ROCS
performance for the same benchmark, using the crystal structure
of a ligand from each chemical class as the ROCS query (Martha
Head, personal communication). For the protein with the most
chemical classes of ligands, PPARD-δ, with ligands from five
chemical classes, ROCS was able to cross all chemical class
boundaries from every starting point. We do not expect shape
similarity methods to find ligands of significantly different
shapes or sizes using a single ligand query, but it has shown
remarkable ability to discover novel chemistries within a shape
The issue of the choice of conformation for a query molecule
is even more difficult. We had assumed the utility of shape
similarity methods derived from the ligand providing a “negative
image” of the active site, into which we could fit new ligands.
This would suggest the need for a bioactive conformation of
the ligand. However, in the Evers20experiment there are no
bioactive conformations available. In this experiment using the
lowest energy conformer of an active molecule proved to be an
effective method for selection of the conformation to be used
as a query. As such, whether an experimental conformation
(which is arbitrary) or a low-energy conformer (which is also
arbitrary in a different way) is used as a query has little effect
on ROCS’s performance. We also have evidence (P. C. D.
Hawkins, unpublished results) that replacing the crystallographic
conformation of a query molecule with a low-energy conformer
from OMEGA has essentially no impact on ROCS’s perfor-
We believe the consistency of ROCS and the lack of any
special parametrization are closely related. One of the hallmarks
of overparametrization is fragility of results. The simple,
relatively parameter-free, shape similarity approaches may help
to avoid such fragility. The shape overlap is one parameter, and
the sum of functional group comparisons (color) is another.
These are combined equally to produce a similarity score. We
suggest that the simplicity of this approach underlies the
generality of application. In fact, in the Evers experiment adding
a special-purpose parameter (a pharmacophore feature) had little
effect on ROCS, whereas it had a dramatic effect on GOLD
It is also worth noting that the nature of the decoys will have
a profound effect on the performance metrics for all the tools.
In the Miteva et al. study there are over 250 thymidine, cytosine,
uridine, and adenosine analogues in the decoy set, while 10
compounds in the same sets of series were designated as actives.
Many of these 250 or so “decoys” are certain to be active to
some degree as TK inhibitors and thus fail to meet the
commonly held criterion for a background or decoy compound,
the criterion being that it is inactive against the target of interest.
In our hands ROCS scored many of these “decoy” compounds
very highly, and we may only assume that the other tools used
in this study did so as well. Therefore, the plots for TK from
this study certainly represent an underestimate of performance
for all of the tools on this data set.
Docking has been shown to perform best when conducted in
a holo enzyme complex due to the manifold conformational
changes that often occur to an apo structure upon ligand binding
(see Erickson et al.42). Docking often suffers a reduction in
performance when conducted on apo structures versus holo
structures.43However, the shape-based approach outlined here
performs well when using an experimental conformation
obtained from a holo structure or, as in the GPCR study, an
arbitrary low-energy conformation of the query molecule. These
two studies indicate that docking may sometimes benefit from
ligand information, which is implicit in the holo structure, more
than ligand-based methods do.
One of the criticisms of the shape-based, ligand-centric
approach has been that it completely ignores protein information,
even when such information exists. One modification to the
standard shape similarity approach would be to postfilter results
based on the in situ alignment and interactions with the protein.
In a study from Wyeth44on virtual screening and lead-hopping
with ROCS, overlays were obtained for candidate molecules to
the query molecule in its crystallographic configuration. Given
that these overlays were in the context of the structure of the
target protein, a force-field cleanup (to eliminate candidate
compounds that gave good overlays but also clashed with the
protein) and energy analysis were applied to the ROCS overlays.
This two-layered approach was found to give good results,
resulting in the identification of an entirely new class of active
compound against the target of interest. Work is underway to
determine if this finding is true generally. While this would be
exciting, we suspect that this approach may equally likely fall
Journal of Medicinal Chemistry, 2007, Vol. 50, No. 1Hawkins et al.
foul of the rigid protein assumption that so often bedevils
Another issue frequently raised in ligand-based studies is
whether the performance of a 3D ligand-based tool is mostly
due to trivial 2D similarity between the query molecule(s) and
the active molecules being searched for. As mentioned above,
the mean and maximum 2D similarities between the query
molecule used in each of the experiments and the active
molecules are tabulated in the Supporting Information. Inspec-
tion of the similarities between the queries and the active
molecules used clearly shows that the success of the shape-
based approach presented here is not due to close structural
similarity between the query and the active molecules. Prelimi-
nary investigation into the comparison between 2D fingerprint
and shape-based similarity shows that they are pleasingly
complementary (data not shown).
In sum, direct comparisons between virtual screening results
from a significant number of docking programs show that a
shape-based ranking method (ROCS) performs at least as well
as and often better than docking. In total, seven different docking
programs were compared to ROCS across 21 different protein
systems (15 unique proteins). Since exactly the same sets of
active compounds and decoys were used in this study as the
published docking studies, the conclusion that a shape-based
approach is competitive seems warranted. ROCS provided
superior performance even when a bioactive conformation of
the ligand was not known. Given the success, speed, ease of
use, predictability, and applicability of ranking using shape and
chemical similarity, we suggest that this approach be given
serious consideration in all projects where high-throughput
virtual screening is warranted.
Acknowledgment. The authors who have made the data sets
from their publications freely available to allow direct com-
parisons such as this are warmly thanked. Particular thanks go
to Prof. Villoutriex for assistance in obtaining the data used in
his study and to Prof. Jain for the interest he took in this work
in its earlier stages. Drs. Martha Head and Gregory Warren
(GSK) are thanked for their assistance with replicating their
docking study with ROCS. Dr. Robert Tolbert is thanked for
his invaluable assistance on the intricacies of Python to one of
the authors (P.C.D.H.), and Dr. Roger Sayle is thanked for his
Supporting Information Available: Tables 1-5 listing the
results of Bissantz, Evers, Cummings, Miteva, and Pham. This
material is available free of charge via the Internet at http://
(1) Kraemer, O.; Hazemann, I.; Podjarny, A. D.; Klebe, G. Virtual
screening for inhibitors of aldose reductase. Proteins: Struct., Funct.,
Bioinf. 2004, 55, 814-822.
(2) Hert, J.; et al. Enhancing the effectiveness of similarity-based virtual
screening using nearest-neighbor information J. Med. Chem. 2005,
(3) Velec, H. F. G.; Gohlke, H.; Klebe, G. DrugScoreCSD-knowledge-
based scoring function derived from small molecule crystal data with
superior recognition rate of near-native ligand poses and better affinity
prediction. J. Med. Chem. 2005, 48, 6296-6306. Kontoyanni, M;
Sokol, G. S.; McClellan, L. M. Evaluation of library ranking efficacy
in virtual screening. J. Comput. Chem. 2005, 26, 11-19.
(4) Ahlstrom, M. M.; Ridderstrom, M.; Luthman, K.; Zamora, I. Virtual
screening and scaffold hopping based on GRID molecular interaction
fields. J. Chem. Inf. Model. 2005, 45, 1313-1320.
(5) Krovat, E. M.; Fruwirth, K. H.; Langer, T. Pharmacophore identifica-
tion, in silico screening and virtual library design for inhibitors of
factor Xa. J. Chem. Inf. Model. 2005, 45, 146-154.
(6) Jorissen, R. N.; Gilson, M. K. Virtual screening of molecular
databases using a support vector machine. J. Chem. Inf. Model. 2005,
(7) Abrahamian, E.; et al. Efficient generation, storage and manipulation
of fully flexible pharmacophore multiplets and their use in 3-D
similarity searching. J. Chem. Inf. Comput. Sci. 2003, 43, 458-466.
(8) Lyne, P. D. Structure-based virtual screening: an overview. Drug
DiscoVery Today 2002, 7, 1042-1047.
(9) Jain, A. N. Virtual screening in lead discovery and optimization. Curr.
Opin. Drug DiscoVery DeV. 2004, 7, 396-404.
(10) Warren, G. L.; et al. A critical assessment of docking programs and
scoring functions. J. Med. Chem. 2006, 49, 5912-5931.
(11) (a) Marsden, P. M.; Puvenendrampillai, D.; Mitchell, J. B. O.
Predicting protein-ligand binding affinities: a low scoring game?
Org. Biomol. Chem. 2004, 2, 231-237. (b) Maiorov, V; Sheridan,
R. P. Enhanced virtual screening by combined use of two docking
methods: Getting the most on a limited budget. J. Chem. Inf. Model.
2005, 45, 1017-1024.
(12) (a) Jenkins, J. L.; Kao, R. Y. T.; Shapiro, R. Virtual screening to
enrich hit lists from high-throughput screening: A case study on
small-molecule inhibitors of angiogenin. Proteins: Struct., Funct.,
Genet. 2003, 50, 81. (b) Forino, M.; et al. Virtual docking approaches
to protein kinase B inhibition. J. Med. Chem. 2005, 48, 2278-2289.
(13) Pavia, A. M.; et al. Inhibitors of dihydropicolinate reductase, a key
enzyme of the diaminopimelate pathway of Mycobacterium tuber-
culosis. Biochim. Biophys. Acta 2001, 1545, 67-75.
(14) Doman, T. N.; et al. Molecular docking and high-throughput screening
for novel inhibitors of protein tyrosine phosphatase-1B. J. Med. Chem.
2002, 45, 2213-2219.
(15) Ward, R. A.; Perkins, T. D. J.; Stafford, J. Structure-based virtual
screening for low molecular weight chemical starting points for
dipetidyl peptidase IV inhibitors. J. Med. Chem. 2005, 48, 6991-
(16) Grant, A. J.; Pickup, B. T. A fast method of molecular shape
comparison: a simple application of a Gaussian description of
molecular shape. J. Comput. Chem. 1996, 17, 1653-1659.
(17) Haigh, J. A.; Pickup, B. T.; Grant, J. A.; Nicholls, A. Small molecule
shape-fingerprints. J. Chem. Inf. Model. 2005, 45, 673-680.
(18) Melani, F.; et al. Field interaction and geometrical overlap: A new
simplex and experimental design based computational procedure for
superposing small ligand molecules. J. Med. Chem. 2003, 46, 1359-
(19) Lemmen, C.; Lengauer, T. Computational methods for the structural
alignment of molecules J. Comput.-Aided Mol. Des. 2000, 14, 199-
(20) Evers, A.; Hessler, G.; Matter, H.; Klabunde, T. Virtual screening
of biogenic amine-binding G-protein coupled receptors: Comparative
evaluation of protein- and ligand-based virtual screening protocols.
J. Med. Chem. 2005, 48, 5448-5460.
(21) Cummings, M. D.; et al. Comparison of automated docking programs
as virtual screening tools. J. Med. Chem. 2005, 48, 962-971.
(22) Pham, T. A.; Jain, A. J. Parameter estimation for scoring protein-
ligand interactions using negative training data. J. Med. Chem. 2006,
(23) Miteva, M. A.; Lee, W. H.; Montes, M. O.; Villoutriex, B. O. Fast
structure-based virtual ligand screening combining FRED, DOCK
and Surflex. J. Med. Chem. 2005, 48, 6012-6021.
(24) Bissantz, C.; Folkers, G.; Rognan, D. Protein-based virtual screening
of chemical databases. 1. Evaluation of different docking/scoring
functions. J. Med. Chem. 2000, 43, 4759-4771.
(25) Cole, J. C.; et al. Comparing protein-ligand docking programs is
difficult. Proteins: Struct., Funct., Bioinf. 2005, 60, 325-331.
(26) OMEGA, OpenEye Scientific Software, Santa Fe, NM. http://
(27) Berman, H. M.; et al. The Protein Databank. Nucleic Acids Res. 2000,
28, 235. http://www.rcsb.org.
(28) OEChem Toolkit, OpenEye Scientific Software, Santa Fe, NM.
(29) ROCS, OpenEye Scientific Software, Santa Fe, NM. http://www.
(30) Mills, J. E. J.; Dean, P. M. Three-dimensional hydrogen-bond
geometry and probability information from a crystal survey J.
Comput.-Aided Mol. Des. 1996, 10, 607-613.
(31) Sheridan, R. P.; et al. Protocols for bridging the peptide to non-
peptide gap in topological similarity searches. J. Chem. Inf. Comput.
Sci. 2001, 41, 1395-1403.
(32) Hanley, J. A.; McNeil, B. J. The meaning and use of the area under
a receiver operating characteristic (ROC) curve. Radiology 1982, 143,
Shape-Matching and DockingJournal of Medicinal Chemistry, 2007, Vol. 50, No. 1 81
(33) Baldi, P.; et al. Assessing the accuracy of prediction algorithms for Download full-text
classification: an overview. Bioinformatics 2000, 16, 412-427.
(34) Tribelleau, N.; et al. Virtual screening workflow development guided
by the “receiver operator characteristic” curve approach. Application
to high-throughput docking on metabotropic glutamate receptor
subtype 4. J. Med. Chem. 2005, 48, 2534-2541.
(35) Renfrey, S.; Featherstone, J. Structural proteomics. Nat. ReV. Drug
DiscoVery 2002, 1, 175-192.
(36) Evers, A.; Klabunde, T. Structure-based drug discovery using GPCR
homology modeling: Successful virtual screening for antagonists of
the Alpha1A adrenergic receptor. J. Med. Chem. 2005, 48, 1088-
(37) MDDR. http://www.mdl.com/products/knowledge/drug_data_report/.
(38) McGann, M. R.; et al. Gaussian docking functions. Biopolymers 2003,
(39) Nicholls, A. N.; MacCuish, N. E.; MacCuish, J. D. Variable selection
and model validation for 2D and 3D descriptors J. Comput.-Aided
Mol. Des. 2004, 18, 451-461.
(40) Bostrom, J; Greenwood, J. R.; Gottfries, J. Assessing the performance
of OMEGA with respect to retrieving bioactive conformations. J.
Mol. Graphics Modell. 2003, 21, 449-457.
(41) Perola, E.; Walters, W. P.; Charifson, P. S. An Analysis of Critical
Factors Affecting Docking and Scoring in Virtual Screening. In
Virtual Screening in Drug DiscoVery, 1st ed.; Alvarez, J., Shoichet,
B., Eds.; Taylor & Francis: Boca Raton, FL, 2005; pp 47-85.
(42) Erickson, J. A.; et al. Lessons in molecular recognition: The effects
of ligand and protein flexibility on molecular docking accuracy. J.
Med. Chem. 2004, 47, 45-53.
(43) McGovern, S. L.; Shoichet, B. K. Information decay in molecular
docking screens against holo, apo and modeled conformations of
enzymes. J. Med. Chem. 2003, 46, 2895-2904.
(44) Rush, T. S., III; Grant, A. J.; Mosyak, L.; Nicholls, A. A shape-
based 3-D scaffold hopping method and its application to a bacterial
protein-protein interaction. J. Med. Chem. 2005, 48, 1489-1494.
Journal of Medicinal Chemistry, 2007, Vol. 50, No. 1Hawkins et al.