Content uploaded by Natasja Brooijmans
Author content
All content in this area was uploaded by Natasja Brooijmans on Jan 29, 2014
Content may be subject to copyright.
ORIGINAL PAPER
Development and validation of a modular, extensible docking
program: DOCK 5
Demetri T. Moustakas Æ P. Therese Lang Æ Scott Pegg Æ Eric Pettersen Æ
Irwin D. Kuntz Æ Natasja Brooijmans Æ Robert C. Rizzo
Received: 12 April 2006 / Accepted: 22 July 2006 / Published online: 6 December 2006
Springer Science+Business Media B.V. 2006
Abstract We report on the development and valida-
tion of a new version of DOCK. The algorithm has been
rewritten in a modular format, which allows for easy
implementation of new scoring functions, sampling
methods and analysis tools. We validated the sampling
algorithm with a test set of 114 protein–ligand com-
plexes. Using an optimized parameter set, we are able to
reproduce the crystal ligand pose to within 2 A
˚
of the
crystal structure for 79% of the test cases using our rigid
ligand docking algorithm with an average run time of
1 min per complex and for 72% of the test cases using
our flexible ligand docking algorithm with an average
run time of 5 min per complex. Finally, we perform an
analysis of the docking failures in the test set and
determine that the sampling algorithm is generally suf-
ficient for the binding pose prediction problem for up to
7 rotatable bonds; i.e. 99% of the rigid ligand docking
cases and 95% of the flexible ligand docking cases are
sampled successfully. We point out that success rates
could be improved through more advanced modeling of
the receptor prior to docking and through improvement
of the force field parameters, particularly for structures
containing metal-based cofactors.
Keywords Automated docking Scoring functions
Structure-based drug design Flexible docking
Binding mode prediction Incremental construction
Validation
Introduction
Transient non-covalent interactions are critical for bio-
logical processes. The sequencing of a variety of ge-
nomes and the development of proteomics techniques
have enabled scientists to study these interactions on the
widest scales [1]. Advances in X-ray crystallography,
nuclear magnetic resonance spectroscopy, and other
experimental structure techniques provide the ability to
study these interactions at an atomic level of detail [2].
One important application of these advances is the de-
sign of small molecules that interact with cellular pro-
cesses to modify biological activity and treat disease.
D. T. Moustakas and P. T. Lang are joint first authors
Electronic Supplementary Material The structure files for
the test set and the optimized input files used to generate this
data can be found at the DOCK web site
(http://dock.compbio.ucsf.edu).
D. T. Moustakas
Joint Graduate Program in Bioengineering, University of
California, San Francisco, 600 16th Street, Genentech Hall,
Box 2240, San Francisco, CA 94143, USA
D. T. Moustakas
Joint Graduate Program in Bioengineering, University of
California, Berkeley, Berkeley, CA, USA
P. T. Lang N. Brooijmans
Graduate Program in Chemistry and Chemical Biology,
University of California, San Francisco, 600 16th Street,
Genentech Hall, Box 2240, San Francisco, CA 94143, USA
S. Pegg E. Pettersen I. D. Kuntz (&)
Department of Pharmaceutical Chemistry, University of
California, San Francisco, 600 16th Street, Genentech Hall,
Box 2240, San Francisco, CA 94143, USA
e-mail: kuntz@cgl.ucsf.edu
R. C. Rizzo
Department of Applied Mathematics and Statistics, Stony
Brook University, Room 1-101, Stony Brook, NY
11794-3600, USA
123
J Comput Aided Mol Des (2006) 20:601–619
DOI 10.1007/s10822-006-9060-4
The drug discovery process typically requires be-
tween 10 years and 15 years from early discovery until
FDA approval [3]. Computational tools—such as vir-
tual screening, homology modeling and cheminfor-
matics—are applied both to facilitate various stages of
research and to create models that explain experi-
mental data [4–6]. Molecular docking, which can
broadly be defined as the prediction of the orientation
of two molecules with respect to one another, is a
computational technique that has been successfully
used in both of these capacities [7]. In drug design
applications, one molecule is typically a protein or
nucleic acid drug target—the receptor—and the other
is a potential ligand. In these applications, docking is
used to identify novel ligands that interact with a bio-
molecular target and to predict the geometric position
(binding mode) of ligands with respect to the target of
interest.
DOCK background
DOCK is one example of a family of molecular
docking packages available, which includes Glide,
FlexX, and GOLD (Table 1)[8–11]. Each of these
programs consists of two key parts: a search algorithm
and a scoring function. The search algorithm samples
both the relative orientations of the two objects as well
as their conformations. It must be thorough enough to
ensure adequate coverage of the binding free energy
landscape in order to find the global minimum of the
scoring function. The scoring function ranks the vari-
ous geometries generated by the search algorithm,
proposing the top-scoring pose as the global minimum.
It must rapidly evaluate receptor–ligand complex sta-
bility with sufficient accuracy such that the global
minimum of the scoring function agrees with experi-
mental data.
The number of degrees of freedom in recep-
tor–ligand interactions is very large, and several
approximations must be made to ensure that the
docking problem is tractable. Many different
approaches, ranging from freezing non-essential mo-
tions to the use of preferred conformations, have been
developed to reduce the number of degrees of freedom
sampled [12]. In the DOCK algorithm, for example,
the receptor is considered to be conformationally rigid,
requiring only the ligand conformational, translational
and rotational degrees of freedom to be sampled dur-
ing complex formation. This assumption is reasonable
in docking applications in which either the receptor
conformation does not change dramatically upon li-
gand binding or in which the aim is to stabilize a par-
ticular receptor conformation.
In order to guide the search for ligand orientations
with respect to the receptor, a negative image of the
active site volume is created by placing spheres on the
solvent accessible surface area of the receptor, thus
restricting the ligand orientational sampling to the
most relevant region on the surface of the receptor
[13]. To sample the internal degrees of freedom of the
ligand, DOCK uses the incremental construction
algorithm, anchor-and-grow, which separates the li-
gand flexibility into two steps [14, 15], (Fig. 1). First,
the largest rigid substructure of the ligand (anchor) is
identified and rigidly oriented in the active site by
matching its heavy atoms centers to the receptor
sphere centers (orientation). The anchor orientations
are evaluated and optimized using the scoring func-
tion and the energy minimizer. The orientations are
then ranked according to their score, spatially clus-
tered by heavy atom root mean squared deviation
(RMSD), and prioritized (pruning). Next, the
remaining flexible portion of the ligand is built onto
the best anchor orientations within the context of the
receptor (grow). It is assumed that the shape of the
binding site will help restrict the sampling of ligand
conformations to those that are most relevant for the
receptor geometry.
Table 1 Summary of scoring functions and sampling algorithms for commonly used docking programs
Method Ligand sampling method
a
Receptor sampling method
a
Scoring function
b
Solvation scoring
c,d
DOCK 4/5 IC SE MM DDD, GB, PB
FlexX/FlexE IC SE ED NA
Glide CE + MC TS MM + ED DS
GOLD GA GA MM + ED NA
a
Sampling methods are defined as Genetic Algorithm (GA), Conformational Expansion (CE), Monte Carlo (MC), incremental
construction (IC), merged target structure ensemble (SE), torsional search (TS)
b
Scoring functions are defined as either empirically derived (ED) or based on molecule mechanics (MM)
c
If the package does not accommodate this option, the symbol NA (Not Available) is used
d
Additional accuracy can be added to the scoring function using implicit solvent models. The most commonly used options are
distance dependent dielectric (DDD), a parameterized desolvation term (DS), generalized Born (GB) and linearized Poisson
Boltzmann (PB)
602 J Comput Aided Mol Des (2006) 20:601–619
123
In order to evaluate a large number of ligand poses
in a reasonable amount of time, approximate scoring
functions must be used. Once again, numerous solu-
tions to this problem have been proposed, including a
variety of empirical and physics-based terms [12].
DOCK uses an energy scoring function based on the
AMBER molecular mechanics force field [14, 16].
Only the interactions between the ligand and protein
are considered, leaving only intermolecular van der
Waals (VDW) and electrostatic components in the
function. Since the receptor is considered to be rigid,
the receptor contribution to the potential energy can
be pre-calculated and stored on a grid [16]. These
approximations enable the program to evaluate large
libraries of small molecules against a receptor in a
reasonable period of time.
This paper describes a new version of the DOCK
program and explores the critical variables that con-
trol its ability to find correct binding modes in a suite
of test problems. Our motivation is to provide a
modular docking package that permits the easy
development of new scoring functions, search algo-
rithms, and analysis tools. Thus, each functional unit
of the DOCK algorithm was implemented as a self-
contained and portable module that interacts with the
user through a well-defined interface (Fig. 2). The
object-oriented language C++ was chosen to allow
each component of the DOCK algorithm to be
implemented as a class, which encapsulates both the
data structures and functions [17]. DOCK 5 incorpo-
rates several new routines, including parallelization of
the algorithm through an external library, modifica-
tion of the ligand structural class to enable greater
user control over sampling, and clustering of the final
results by root mean square deviation. The implica-
tions of these additions will be discussed in this
paper. Additional scoring functions and alternate
sampling techniques have been implemented as
well and will be discussed in future papers (http://
dock.compbio.ucsf.edu).
Previous studies have examined the scoring function
and the matching algorithm of DOCK in detail ([14]
and equations 1–6 in [16]). In this paper, we pay par-
ticular attention to the robustness of the anchor-and-
grow portion of the DOCK algorithm. We seek to
maximize the success of complex structure prediction
by independently optimizing the various steps in the
anchor-and-grow algorithm. In the process, we also
quantify and bound the errors for cases in which flex-
ible docking fails and provide direction for potential
areas of improvement.
Fig. 1 The ‘‘anchor-and-grow’’ conformational search algo-
rithm. The algorithm performs the following steps: (1) DOCK
perceives the molecule’s rotatable bonds, which it uses to
identify an anchor segment and overlapping rigid layer segments.
(2) Rigid docking is used to generate multiple poses of the
anchor within the receptor. (3) The first layer atoms are added to
each anchor pose, and multiple conformations of the layer 1
atoms are generated. An energy score within the context of the
receptor is computed for each conformation. (4) The partially
grown conformations are ranked by their score and are spatially
clustered. The least energetically favorable and spatially diverse
conformations are discarded. (5) The next rigid layer is added to
each remaining conformation, generating a new set of confor-
mations. (6) Once all layers have been added, the set of
completely grown conformations and orientations is returned
J Comput Aided Mol Des (2006) 20:601–619 603
123
Overview of test set
The validation of any software program requires
careful testing of all aspects of the algorithm and
assessment of its utility in all anticipated applications
of the software. Molecular docking is commonly used
in several modes, namely ligand binding mode pre-
diction, virtual screening, and prioritization of a set
of related compounds based on their affinity. How-
ever, predicting the correct binding mode of a li-
gand–receptor complex is a requisite step for the
successful comparison of different ligands and
therefore will be the focus of this paper. It is
important to note, however, that predicting binding
orientations is not the only metric for the accuracy
and utility of docking algorithms. Optimizing DOCK
for applications, including ranking libraries of small
molecules and calculating absolute free energies
of binding, will be addressed in other papers
(http://dock.compbio.ucsf.edu).
Large-scale validation of docking algorithms was
long hampered by the lack of a large number of high
quality protein–ligand complex crystal structures.
Thanks to advances in automation in molecular biol-
ogy and crystallography, the number of structures in
the Protein Data Bank (PDB) continues to grow at a
rapid pace [18]. The developers of GOLD were first to
test their program on a large number of available
structures [19]. Their test set was compiled using a
number of criteria to select candidate protein–ligand
complex structures. The protein must be of pharma-
cological interest and the ligands must be drug-like. In
addition, complexes were chosen that exhibited inter-
esting and unusual interactions between the ligand and
the protein. The final set of 100 (more recently ex-
panded to 134) protein–ligand complexes has served as
the basis for other, larger test sets [11, 20–22].
More recently, the CCDC/Astex set compiled 305
protein–ligand complex structures by expanding the
original GOLD test set [22]. However, the authors
note that many of the new entries contain larger li-
gands that have more rotatable bonds, making this set
less drug-like. The crystal structures in the CCDC/
Astex set were evaluated for crystallographic errors
and inconsistencies, yielding a ‘‘clean’’ set of 224 pro-
tein–ligand complexes. To create the test set for the
DOCK validation studies, we filtered out 84 complexes
with eight or more rotatable ligand bonds. In addition,
several of the complexes had properties that we felt
made them inappropriate for a validation set. These
issues included ligands that were covalently bound to
the receptor (PDB code 1ASE), ligands with missing
Fig. 2 The major DOCK 5 classes and their interconnections.
The bold arrows denote the connections between the classes that
implement the DOCK sampling algorithm. The path traced by
the arrows illustrates the sequence of operations performed upon
a ligand molecule during docking. The bold lines (without
arrowheads) denote functional connections between classes.
These connections allow one class to call functions implemented
in another. This diagram demonstrates that the classes imple-
menting the DOCK sampling methods are heavily connected to
a layer of classes that implement the physics engine: the force
field, the scoring functions, and the energy minimizers. The thin
lines denote hierarchical relationships between a master class
and modular subclasses. These hierarchical arrangements allow
new functional classes (scoring functions, energy minimizers,
etc.) to be plugged into the existing DOCK algorithm in a
modular fashion
604 J Comput Aided Mol Des (2006) 20:601–619
123
electron density (PDB code 1EED), and known se-
quence misregistry in the receptor (PDB code 3HVT).
Ligands with vanadium that required VDW types in
which we were not completely confident were also re-
moved. The final test set contained 114 drug-like
complexes (see Methods, Table 2).
Methods
DOCK 4 to DOCK 5 conversion
The new DOCK rigid body orienting code was written
as a direct implementation of the isomorphous sub-
graph matching method of Kuhl et al. [23]. All receptor
sphere pairs and atom center pairs are considered for
inclusion in a matching clique. This is more computa-
tionally demanding than the clique matching algorithm
implemented in previous versions of DOCK that used
a distance binning algorithm to restrict the clique
search, in which pairs of spheres and atom centers were
binned by distance. Only sphere pairs and center pairs
that were within the same distance bin were considered
as potential matches [14]. The new DOCK clique
matching implementation avoids bin boundaries that
prevent some receptor sphere and ligand atom pairs
from matching, and, as a result, it can find good mat-
ches missed by previous versions of DOCK. The rigid
body rotation code was also corrected to avoid a sin-
gularity that occurred if the spheres in the match lay
within the same plane. Both of these changes improved
orientational sampling.
The anchor-and-grow algorithm in the new version
of DOCK was also modified to prevent premature
pruning of the growth tree. The DOCK 5 anchor-and-
grow code was completely rewritten with several dif-
ferences in the implementations between DOCK 4 and
5. The anchor-and-grow implementation in DOCK 5
fixed a series of bugs that caused some branches of the
search to be pruned when they should have been pre-
served for the next round of growth. The mechanism of
minimization of partially grown conformers was also
changed to allow the entire partial conformer to move,
instead of just the latest layer, enabling more accurate
ranking and pruning of the partially grown conformers.
In addition, the simplex minimizer was re-coded
based on the original Nelder and Mead algorithm [24].
The new minimizer implementation consistently found
lower energy minima when using the same set of 1,000
ligand orientations in a receptor, indicating that it was
performing better than the previous version (data not
shown). In addition, we changed the mechanism of
minimization of partially grown ligand conformers to
allow all atoms in the partial conformer to be mini-
mized, rather than only the outermost layer of atoms.
These changes may explain why DOCK 4 performs
more poorly when run with the DOCK 5 optimized
parameters (see below).
The final version of the new DOCK code, including
all functions described below and all bug fixes, was
posted to the DOCK web site as version 5.4.0 (http://
dock.compbio.ucsf.edu). All experiments performed
with the new implementation of DOCK used this
version and will be referred to as DOCK 5 for conve-
nience. All experiments performed with the previous
version of DOCK used version 4.0.1 and will be re-
ferred to as DOCK 4.
Conversion of the DOCK codebase from C to C++
The design of the new DOCK 5 architecture balances
the speed of the code, or computational performance,
against its modularity and extensibility. The code was
developed using ANSI C++ to ensure portability across
multiple platforms [17]. The only external library used
by DOCK 5 is MPICH for parallel processing [25]. To
enable easy modification or replacement of DOCK 5
algorithm components, the DOCK 5 class structure
was designed so that there are classes for each major
DOCK algorithm function, and these classes interface
with each other by passing instances of the DOCK 5
molecule class. Within the major functions, there are
two layers of classes: those that implement the ligand
sampling functions—rigid orienting, conformational
searching, and minimizing—and those that implement
Table 2 Complexes used in the test set (total of 114 complexes)
Protein data bank identifier
1A28 1COM 1FLR 1OKL 1TYL 2MCP
1A6W 1COY 1HAK 1PBD 1UKZ 2PCP
1A9U 1CPS 1HDC 1PDZ 1ULB 2PHH
1ABE 1D3H 1HSL 1PHD 1WAP 2PK4
1ABF 1D4P 1HYT 1PHG 1XID 2TMN
1ACJ 1DBB 1IMB 1PTV 1XIE 2YPI
1ACM 1DBJ 1IVB 1QCF 1YDR 3CPA
1ACO 1DG5 1LAH 1QPE 2AAD 3ERD
1AI5 1DID 1LCP 1QPQ 2ACK 3GPB
1AOE 1DOG 1LDM 1RNT 2ADA 3HVT
1AQW 1DR1 1LST 1ROB 2AK3 4AAH
1AZM 1DWB 1LYL 1RT2 2CHT 4COX
1BYG 1EBG 1MDR 1SNC 2CMD 4CTS
1C5C 1ETT 1MLD 1SRJ 2CPP 4FBP
1C5X 1F0R 1MRG 1TDB 2CTC 4LBD
1C83 1F0S 1MRK 1TNG 2DBL 5ABP
1CBX 1F3D 1MUP 1TNH 2GBP 5CPP
1CIL 1FGI 1NGP 1TNI 2H4N 6RNT
1CKP 1FKI 1NIS 1TNL 2LGS 7TIM
J Comput Aided Mol Des (2006) 20:601–619 605
123
the underlying physics engine—the force field defini-
tions and the scoring functions. The sampling classes
are applied sequentially to the ligand molecule; the
physics engine classes are utilized by the sampling
classes to score the ligand–receptor interaction after
each step.
As a specific example of modularity, the DOCK 5
scoring functions are implemented as a master score
class with five scoring function subclasses. The master
score class acts as an interface to the scoring subclasses,
enabling the user to designate primary and secondary
scoring functions at runtime. This design was chosen
because the individual scoring functions were best
implemented as individual classes; they each require
different input and use different internal data struc-
tures. While they could have been implemented into
one large scoring class, the result would have been
quite large and disjoint. This solution was also applied
to the ligand conformational search, energy minimi-
zation and post-docking analysis classes.
The DOCK 5 molecule class was designed to con-
tain the minimum information required to specify a
three-dimensional ligand conformation (atom coordi-
nates, bond connectivity, atom partial charges, atom
types and bond types) to minimize the memory re-
quired to store a molecule, allowing large arrays of
molecules to be stored in RAM. Standard C-style ar-
rays were used to store the molecular data to maximize
the speed of accessing this information.
Test set preparation
The proteins and ligands were extracted from the PDB
files, which were downloaded from the PDB website
(www.rcsb.org, Table 2). The ligands were assigned
atom types and bond types manually, and hydrogens
were added using Sybyl [26]. Subsequently, AM1-BCC
partial electrostatic charges were calculated using the
Antechamber package distributed with Amber 8 [27,
28]. The number of rotatable bonds of each of the li-
gands was measured using DOCK, and ligands
with > 7 rotatable bonds were eliminated from the test
set. We choose seven or fewer bonds to give a rea-
sonable representation of DOCK’s performance using
compounds similar to those of most interest in drug
discovery [29–31]. The final test set that was used
consisted of 114 non-covalent protein–ligand com-
plexes [32] (Table 2).
For the proteins, we removed all waters, covalently
linked sugars, sulfates, and halogens that were not
part of the ligand. Co-factors, such as heme, ATP,
and NADPH, were kept, atom and bond types were
assigned manually, and Gasteiger–Hu
¨
ckel partial
electrostatic charges were calculated using the
‘‘Compute’’ module in Sybyl [26, 33, 34]. Ions, such as
calcium and zinc, were considered to be part of the
protein and the correct charge was assigned manually.
Different VDW parameters for zinc were used
depending on the coordination state of the zinc atom
in the protein–ligand complex (Table 3). Hydrogens
were added to the protein residues using the ‘‘Bio-
polymer’’ module in Sybyl, as were AMBER partial
charges and VDW parameters [26, 37]. No additional
optimization of the protein structure was carried out
at this point.
The GRID accessory program of DOCK was used
to pre-calculate scoring function potential grids [16].
All parameters were set to default parameters, except
for the ‘‘energy_cutoff_distance,’’ which was set to
9,999, resulting in the inclusion of all protein atoms in
the energy calculation. For matching, the dms program
was used to generate a molecular surface for each
receptor [38]. The SPHGEN accessory program of
DOCK was used to create a negative image of the
surface using spheres [39, 40]. For the purpose of this
validation study, a general procedure was established
to generate a sphere cluster for every protein in the test
set. In this procedure, we select all the spheres found
within 10 A
˚
of any ligand atom. The receptor box
delimiting the active side was calculated with the
accessory program SHOWBOX using the sphere set
with an additional 5 A
˚
boundary. We have explored
additional box sizes ranging from 1 A
˚
to 9 A
˚
padding
and found that there is little sensitivity to the exact
padding amount (i.e. success rate for rigid ligand
docking of 80 ± 1%, time increase 10% with padding
size increase, and an average test set energy of -
50 ± 0.1 DOCK units). The final procedure creates
sphere sets with an average of 101 docking spheres and
boxes of ~20 A
˚
3
. These receptor sphere sets are larger
than what one would typically use in most docking
applications. This adds stringency to our testing of
DOCK 5 by increasing the orientational and transla-
tional space that it must search.
Table 3 Zinc VDW parameters used to generate grids
Tetra-coordinated Zinc
a
Radius 1.700 A
˚
Well depth 0.067 kcal/mol
Penta-coordinated Zinc
b
Radius 1.100 A
˚
Well depth 0.0125 kcal/mol
a
Parameters used for receptors with tetra coordinated zinc ions
[35]
b
Parameters used for receptors with penta coordinated zinc ions
[36]
606 J Comput Aided Mol Des (2006) 20:601–619
123
Optimized hydrogen locations for test set receptors
To assess the effect of hydrogen placements on dock-
ing outcomes, we also optimized the hydrogen atom
placement and hydrogen-bonding network for the
receptor using the ‘‘Dock Prep’’ module in Chimera
[41]. In this module, the hybridization states of the
non-hydrogen atoms of a PDB structure are deter-
mined by an enhanced version of the IDATM atom-
typing algorithm [42]. Then, all hydrogens that can be
unambiguously positioned are added to the file. To
assist in positioning ambiguous hydrogens, hydrogen-
bonding interactions are examined. The definitions of
hydrogen-bonding donors and acceptors as well as
hydrogen-bonding angle and distance criteria are based
on the values found in Mills and Dean [43]. Relevant
hydrogen bonds (H-bonds) are examined from shortest
to longest, with satisfaction of shorter bonds having
priority. For H-bonds where it is unclear which end is
acting as the donor (e.g. water–water), use of that bond
is postponed until either end is resolved further,
though any lower-priority bonds that conflict geomet-
rically with the postponed bond are eliminated from
consideration at that time. If neither end is resolved by
other interactions, the ambiguity is decided arbitrarily.
Should examination of H-bond interactions not com-
pletely determine the positions of all of the hydrogens
bound to a heavy atom, they are positioned to first
satisfy potential H-bond interactions, then any
remaining hydrogens are positioned to avoid steric
clashes with other atoms. For histidine residues, nor-
mally one nitrogen will be protonated (chosen based
on H-bond/steric considerations); however if both ring
nitrogens are H-bond donors, they will both be pro-
tonated.
Selection of active site waters
All waters within 3 A
˚
RMSD of any ligand heavy atom
were selected. These waters were included as part of
the receptor. The new receptor–water complexes were
then subjected to the same hydrogen bonding optimi-
zation as above.
DOCK parameter optimization
To characterize the performance of DOCK 5 in
regenerating known complex structures, we explored
the optimum parameters for use with rigid and flexible
ligand docking strategies (see Appendix 1). Unless
otherwise stated, all docking experiments were carried
out on 2.2 GHz dual processor Opteron 828s running
Linux Fedora Core 3. The code was compiled using
open-source GNU compilers (http://www.gnu.org).
The optimized parameters have been implemented as
the defaults. We note that our primary criterion for
optimization was success in finding the proper ligand
geometry and not the CPU time required per com-
pound. Unless otherwise stated, these parameters were
used for all experiments in this paper.
Greedy clustering of conformational ensemble
The greedy clustering algorithm is designed to elimi-
nate redundant ligand orientations from consideration.
DOCK generates a set of ligand orientations that are
ranked by the scoring function. The RMSD between
each ligand orientation in the list is calculated. If the
RMSD between two ligand orientations falls within the
clustering threshold, the second orientation is assigned
to a cluster with the first. The first ligand orientation is
selected and compared to all subsequent unclustered
orientations in the list; this process is repeated until the
last unclustered orientation has been selected. Once
the entire list has been processed, only the best scoring
ligand pose in each cluster, designated as the cluster
head, is retained.
Evaluation of MPI functionality
Parallel processing is fully integrated into the DOCK
calculation. The DOCK program starts a single master
node and a set of processing nodes. The master node
performs file processing and molecule input/output,
whereas the processing nodes perform the actual
docking calculations. If the number of processors is set
to 1, the code defaults to non-MPI behavior. As a re-
sult of this configuration, there will be minimal differ-
ence in performance between 1 and 2 processors.
Improved performance will only become evident with
more than two nodes. It should be emphasized that the
primary benefit in using DOCK 5 in parallel mode is to
reduce bookkeeping tasks associated with manually
splitting up a database into multiple chunks, which
then must be submitted to different processors indi-
vidually. DOCK 5 automatically partitions out subsets
of a database to various nodes, collates and ranks the
final results, and takes care of all intermediate book-
keeping.
To gauge the performance of parallelization of the
DOCK 5 algorithm, two small subsets of the NCI
database from the ZINC database were constructed
[25, 44]. The two subsets, one containing 500 and the
other 1,000 small molecules, were filtered to have £5
and £14 rotatable bonds, respectively. The receptor
used as a target for this study was HIV-1 reverse
J Comput Aided Mol Des (2006) 20:601–619 607
123
transcriptase in complex with nevirapine (PDB code
1VRT). Because the receptor was not part of the test
set, nevirapine was flexibly redocked using the opti-
mized parameters, which yielded a ligand orientation
0.28 A
˚
RMSD from the crystal structure orientation.
In addition, a library consisting of 1,000 copies of ne-
viripine was generated to remove dependence on the
order and size of the compound library. All parallel-
ization study calculations were executed at the Com-
putational Science Center at Brookhaven National
Laboratory (http://www.bnl.gov/csc) on a cluster con-
sisting of 34 nodes with dual 3.2 GHz Xeon processors
running Linux. Tests were performed using between 2
and 68 nodes. The code was compiled using open-
source GNU compilers and MPI software mpich ver-
sion 1.2.7 from Argonne National Laboratory (http://
wwwunix.mcs.anl.gov/mpi/mpich).
Results
We first consider the results of rigidly docking ligands,
which used a conformation taken directly from the
complex crystal structure, to the complex crystal
structure conformation of the receptor. We then pres-
ent the results of flexible ligand docking tests. In each
case, we consider (a) the overall performance of each
sampling algorithm, (b) the ability of each algorithm to
reproduce the crystal ligand orientation as the top-
scoring pose, (c) the effect of the initial ligand con-
formation on the performance of the algorithm, (d) any
additional information contained in the set of all
sampled ligand orientations, and (e) the ability to ex-
tract additional information by clustering docking re-
sults. We also compare the performance of DOCK 5 to
equivalent DOCK 4 experiments. Finally, we analyze
the cases in which DOCK 5 fails to reproduce the
crystal structure and propose some directions for
improvement of both the DOCK algorithm and our
test set preparation method.
Rigid ligand docking
Overall performance
Unless otherwise noted, all experiments described in
this section involved rigid docking of the complex
crystal structure ligand conformation to the receptor
complex crystal structure. For each case in the test set,
the heavy atom RMSD between the top-scoring
docked ligand pose and the complex crystal structure
ligand pose was evaluated. A DOCK 5 run was con-
sidered to be successful for cases in which the RMSD
between for the top-scoring ligand orientation and the
crystal ligand orientation was less than 2.0 A
˚
.DOCK5
selects the correct pose as the lowest energy structure
for 79% (90/114) of the test cases using the rigid
docking protocol with an average time of 55 s per
complex.
Dependence on ligand conformation
An ensemble of ligand conformations was generated
using the anchor-and-grow algorithm to apply changes
of each of the ligand’s rotatable bonds. This expan-
sion generated a conformation ensemble for each li-
gand that covered all torsional parameters that
DOCK samples. Each generated conformation was
rigidly docked to the receptor, and the results from all
the dockings were binned according to the magnitude
of the ligand’s conformational perturbation (Fig. 3a).
The curve shows dramatic and continual decrease in
the success rate as the perturbation magnitude in-
creases with little success for any ligand conforma-
tions greater than 0.5 A
˚
heavy atom RMSD away
from the crystal conformation. Therefore, any con-
formation generation method must generate ligand
conformations within 0.5 A
˚
heavy atom RMSD of the
crystal conformation for rigid docking to have a rea-
sonable chance to succeed.
Analysis of total orientational ensemble
To this point, we have disregarded ‘‘near misses,’’
which we define as any generated orientations within
2A
˚
RMSD from the crystal structure that are close to
the top of the ranked conformation list, but are not the
best scoring poses. We can examine the remaining
poses either by including all poses that differ by a fixed
energy unit from the most favorable geometry or by
including those that differ by a fixed number of ranked
poses from the most favorable energy. In order to
quantify the extent of these partial successes, all gen-
erated ligand poses for each test case were preserved
and sorted by their energy scores.
An energy gap is defined as the difference between
the DOCK score of the top scoring ligand orientation
and the score of a ligand ranked further down the list.
Considering all docked ligand orientations with an
energy gap of 2.5 DOCK units—an average of five li-
gand orientations—increases the rigid ligand docking
success rate to 90% for the entire test set, while an
average of 50 orientations increase the rigid docking
success rate to 99% (Fig. 4a, b). These results indicate
that the orienting method samples near-crystal ligand
orientations well, but the current energy scoring func-
608 J Comput Aided Mol Des (2006) 20:601–619
123
tion cannot discriminate well between the top-ranked
orientations.
Geometric clustering of poses
Each ligand conformational ensemble was spatially
clustered according to inter-pose RMSD values (see
Methods section for algorithm details). After examin-
ing a range of potential cut-offs, an optimal value of
1.0 A
˚
was chosen (Fig. 5). Using this clustering
threshold, only 15 clusterheads are required to achieve
a success rate of 99%, compared with the top 50
ranked unclustered orientations. This result is encour-
aging, suggesting that the clustering helps sort through
the conformers efficiently.
Flexible ligand docking
Overall performance
Unless otherwise noted, all experiments described in
this section involved flexible docking of the ligand to
the receptor complex crystal structure. As with the ri-
gid docking tests, the heavy atom RMSD between the
a) b)
Fig. 3 (a) Rigid docking
success rates (n)—as
calculated by any
conformation being within
2A
˚
heavy atom RMSD of the
complex crystal
orientation—shown as a
function of the ligand internal
conformation perturbation
magnitude (RMSD). (b)
Flexible growth success rates
(S)—as calculated by any
conformation being within
2A
˚
heavy atom RMSD of the
complex crystal
orientation—shown as a
function of the magnitude of
the anchor perturbation
(RMSD)
a) b)
Fig. 4 (a) The rigid (n) and
flexible (S) docking success
rate as a function of the
DOCK score energy gap
(kcal/mol) for all conformers
generated. (b) The rigid and
flexible docking success rate
as a function of the number of
ranked conformers examined
J Comput Aided Mol Des (2006) 20:601–619 609
123
top-scoring docked ligand pose and the complex crystal
structure ligand pose was evaluated for each complex
in the test set. The success rate over the entire test set
using the optimized flexible ligand anchor-and-grow
protocol was 72% (82/114) with an average time of
314 s per complex.
Dependence on anchor position
The anchor-and-grow algorithm belongs to the set of
incremental construction algorithms for searching li-
gand conformational space [14, 15]. It uses a rigid
docking step for the ‘‘anchors’’ to identify likely anchor
positions (anchor orienting), and a torsion angle search
step to generate ligand conformations rooted at the
previously identified anchor positions (flexible growth).
In order for flexible docking to succeed, both of these
individual steps must be successful.
To measure the dependence of success rate on the
precision of the anchor location, the crystal position of
the anchor for each complex in the test set was per-
turbed randomly from 0 A
˚
to more than 10 A
˚
. Each
perturbed anchor position was then considered as the
starting point for flexible growth (Fig. 3b). With the
anchor starting less than 0.5 A
˚
heavy atom RMSD
from the crystal orientation, the growth algorithm can
find the experimental orientation 99% of the time.
However, the results demonstrate a rapid decrease in
success rate as the anchor is moved further away from
its crystal structure position, decreasing to 76% at
1.0 A
˚
perturbation down to 54% at 2.0 A
˚
. These data
imply that if the flexible ligand docking algorithm can
place the anchor within 0.5 A
˚
heavy atom RMSD of
the crystal anchor position, DOCK 5 has a very high
probability of successfully predicting the full binding
mode correctly.
Analysis of total conformational ensemble
We examined the entire ensemble of conformers gen-
erated by flexible docking, as we described previously
in the rigid ligand docking analysis. Considering all
docked ligand conformations with a 2.5 DOCK unit
energy gap—an average of five ligand orienta-
tions—increases the success rate to 82%, while an
average of 100 orientations increasing the success rate
to 95% (Fig. 4a, b). Again, these results indicate that
the sampling density produced by the optimized
parameters is quite high, but there is little discrimina-
tion between very similar poses by the current scoring
function.
Geometric clustering of poses
As with the rigid ligand docking tests, each confor-
mational ensemble was spatially clustered according to
interpose RMSD (see Methods section for algorithm
details). A clustering threshold of 1.0 A
˚
, as determined
in the rigid docking section, was used (Fig. 5). Using
this clustering threshold, only 50 clusterheads must be
examined to reach a success rate of 95% as compared
to 100 purely ranked orientations. Once again, this
result is encouraging, as it requires a small number of
ligand poses to be retained for rescoring with more
a) b)
Fig. 5 The rigid (filled) and
flexible (open) docking
success rate as a function of
the number of cluster heads
examined. Clusters with
heavy atom RMSD cutoffs of
1.0 A
˚
(
d), 3.0 A
˚
(m), and
5.0 A
˚
(r) were compared
610 J Comput Aided Mol Des (2006) 20:601–619
123
advanced scoring functions that are better at discrimi-
nating between very similar ligand poses.
Comparison to DOCK 4
Using the optimized DOCK 5 parameters, we per-
formed the same rigid and flexible ligand docking
experiments on the entire test set using the last avail-
able version of DOCK 4. The performance of the
current implementation of DOCK 5 compared favor-
ably with the DOCK 4 performance (Table 4). We
attribute the improved accuracy in performance to
improvements outlined in the Methods Section. How-
ever, when comparing the speed of docking experi-
ments between DOCK 4 and DOCK 5, DOCK 4 is
fivefold faster for rigid docking and 30-fold faster for
flexible ligand docking than DOCK 5 (Table 5). We
attribute this increased calculation time to extra stages
of minimization and sampling in DOCK 5, as well as
additional overhead necessary to preserve the modu-
larity of the code (see Methods).
Comparison to other docking methods
Developers of Glide, GOLD and FlexX have also
evaluated their methods using similar test sets and
made some of their analyses available [9, 45, 46]. Based
on this data, we note that DOCK’s flexible docking
success rate of 70% is comparable to Glide’s and
FlexX’s success rates of 82% and 61%, respectively
(Table 6). Unfortunately, GOLD has not posted the
results for the entire CCDC/Astex test set, so a com-
plete comparison could not be made. However, for the
subset of the test set they did report, DOCK’s success
rate of 67% is once again reasonable as compared to the
success rate of 77% for GOLD, considering that the
DOCK scoring function does not use either empirically
weighted parameters or adjustable parameters.
Analysis of successes and failures of docking
protocols
Docking failures can be categorized into two catego-
ries: sampling (soft) and scoring (hard) failures [47].
For scoring failures, an orientation near the crystal
structure was sampled in the course of the DOCK run,
but the scoring function failed to rank it at the top of
the list. A sampling failure indicates that the DOCK
run failed to sample any orientations within 2 A
˚
RMSD of the crystal structure. The major caveat of
this classification scheme is the assumption that the
model of both the receptor and the ligand, including
the VDW parameters, electrostatics, and hydrogen
orientations and protonation states, reflect those that
occur in the experimental structure [48]. Here, we
analyze the flexible docking ligand failures within the
sampling-scoring classification scheme.
Failures resulting from receptor modeling/structural
problems
The original CCDC/Astex test set was filtered for
experimental errors using a variety of metrics [22]. We
plotted the flexible ligand success rate as a function of
various metrics of the quality of the X-ray structures to
determine if the selection criteria were appropriate for
testing the DOCK algorithm (Fig. 6). There appears to
be at best a weak correlation between the RMSD of
the best scoring DOCK pose and either crystal reso-
lution or b-factor of active site or backbone atoms,
indicating that the cut-offs chosen for the original set
were reasonable for docking purposes.
We next explored whether specific atom types caused
problems with the DOCK force field terms by corre-
lating the test set success rate with the presence and type
of active site cofactor (Table 7). The only clear problem
involved metal ions in the receptor. These structures
showed a much lower success rate, accounting for nearly
half of both the rigid and flexible ligand docking failures.
However, there still are a number of failures in the
portion of the test set without cofactors in the active site
that require further characterization. Unless otherwise
mentioned, all studies below were performed on this
subset, referred to as the Cofactor Free (CF) subset.
Table 4 Success based on DOCK version (see Methods)
DOCK version Rigid ligand Flexible ligand
4.0.1 71.9% 42.1%
5.4.0 79.0% 71.9%
Table 5 Average length of time in seconds for docking calcu-
lation using the optimized parameter set (see Appendix 1)
Average Minimum Maximum
DOCK 4 rigid lig 10.9 ± 12.1 0.99 66.8
DOCK 4 flexible lig 7.1 ± 6.04 0.44 33.5
DOCK 5 rigid lig 55.4 ± 37.5 6.0 198.0
DOCK 5 flexible lig 314.7 ± 449.8 2.0 2638.0
Table 6 Comparison of DOCK success rates to other docking
programs for flexible ligand docking
Program No. of complexes Success DOCK success
GOLD 43 77% 67%
Glide 71 82% 70%
FlexX 71 61% 70%
J Comput Aided Mol Des (2006) 20:601–619 611
123
For all members of the test set, the experimental
resolution of the crystal structures was too poor to
identify hydrogen atom locations. We originally mod-
eled the hydrogen atom positions using a rule-based
method. To test this scheme, we applied a more ad-
vanced hydrogen addition procedure that accounted
for steric clashes and hydrogen-bonding networks to
the CF subset (see Methods). As a follow-up, we as-
sumed all crystallographically bound waters found
within 3 A
˚
of any ligand heavy atom were critical for
binding and included them in the receptor model as
well. We found that both of these procedures improved
the flexible ligand docking success rate (Table 8).
Failures resulting from ligand flexibility
In addition to the selection criteria imposed on the
original test set, we also filtered out complexes in
which the ligand had greater than seven rotatable
bonds (see Methods). We reexamined this choice on
the CF subset by plotting the rigid and flexible ligand
docking success rate as a function of the number of
flexible bonds (Fig. 7). As expected, the results show a
decrease in the success rate with increasing ligand size,
but with no dramatic drop-off.
Sampling versus scoring failures
We now return to classification of DOCK failures
based on scoring and sampling classifications [47].
First, we examined the test set failure cases with active
site cofactors (Table 9). Within this set, nine examples
were scoring failures for both rigid and flexible ligand
docking, indicating that new VDW and electrostatic
parameters need to be developed for magnesium,
heme groups, and some coordination states of zinc. In
addition, there were three flexible ligand scoring fail-
ures that were rigid successes, thus suggesting that the
flexible algorithm was able to identify additional ori-
entations with better scores than the experimental li-
gand orientation. Only two flexible ligand docking
cases were sampling failures. We expected flexible li-
gand docking sampling failures due to the increased
ligand degrees of freedom compared with rigid ligand
docking, but it does not appear to be a severe problem
in this test set containing ligands with less than eight
Fig. 6 Correlation of flexible
ligand success (filled) and
failure (striped) rates with
crystallographic resolution
(A
˚
) and experimental
B-factor (A
˚
2
). For active site
B-factors, the active site was
defined as any atom within
9A
˚
of the experimental
ligand orientation
Table 7 Success as function of active site cofactor
Total count Rigid
success
Flexible
success
Entire test set 114 79.0% 71.9%
CF subset 76 81.6% 76.3%
Active site cofactor 38 73.7% 63.2%
Active site metal cofactor 28 64.3% 50.0%
Table 8 Flexible ligand success as function of CF test set
preparation (total of 76 complexes)
Test set preparation technique Success
Standard 76.3%
Hydrogen optimization 78.9%
Active site waters + hydrogen optimization 80.3%
612 J Comput Aided Mol Des (2006) 20:601–619
123
rotatable bonds. Finally, one of the rigid ligand dock-
ing scoring failures was a flexible ligand success. In this
case, there was a large VDW clash between one of the
ligand atoms and the receptor. The anchor-and-grow
algorithm was able to build the ligand in the active site
to avoid this clash, which the rigid ligand docking
algorithm could not accommodate.
We repeated this analysis with the CF subset
(Table 10). Here, there was one rigid ligand docking
sampling failure, which also failed for flexible ligand
docking. Upon closer examination of the receptor site,
a residue making critical interactions with the ligand
was not resolved in the experimental complex structure
(PDB code 1A6W). We anticipate that there may not
be enough contacts to correctly place the molecule.
Seven examples were scoring failures for both rigid and
flexible ligand docking. In this subset, though, we
cannot attribute the failure to unusual atom types,
indicating that the scoring function is incorrectly
modeling some portion of the energy landscape. There
were also seven scoring failures for flexible ligand
docking that were successes for rigid ligand docking,
once again suggesting that the flexible docking algo-
rithm identified additional orientations that scored
better than the experimental orientation.
As in the cofactor set above, there were only three
additional flexible ligand docking sampling failures.
One of these was also a scoring failure in rigid ligand
docking, implying that this failure case may actually be
due to a combination of both sampling and scoring
factors. The remaining two flexible ligand docking
sampling failures once again indicate that the flexible
algorithm was able to identify alternative orientations
that scored better than the crystal complex orientation.
Finally, five rigid ligand docking scoring failures were
flexible ligand dockings successes, signifying that the
b)
a)
Fig. 7 Rigid and flexible
docking success (filled) and
failure (striped) rates as a
function of the number of
rotatable bonds in each ligand
in CF test set
Table 9 Comparison of success and failure cases of both rigid
and flexible docking for complexes in test set with cofactors in
active site (total of 36 complexes)
Rigid
sampling
failure
Rigid
scoring
failure
Rigid
success
Flexible sampling failure 0 0 2
Flexible scoring failure 0 9 3
Flexible success 0 1 23
Table 10 Comparison of success and failure cases of both rigid
and flexible docking for complexes in CF subset (total of 76
complexes)
Rigid
sampling
failure
Rigid
scoring
failure
Rigid
success
Flexible sampling failure 1 1 2
Flexible scoring failure 0 7 7
Flexible success 0 5 53
J Comput Aided Mol Des (2006) 20:601–619 613
123
flexible ligand docking algorithm is able to compensate
for intermolecular clashes in the active site of the
experimental structure that the rigid ligand algorithm
simple cannot accommodate (data not shown).
Analysis of DOCK score for docking protocols
To analyze the ability of DOCK to reproduce the li-
gand–receptor interaction energy as measured by the
DOCK scoring function, we plotted the score from the
top-ranking pose for both rigid and flexible ligand
docking that were successful against the DOCK score
of the complex crystal structure (Fig. 8a, b). Each
crystal structure ligand was minimized with 1,000 steps
of the DOCK simplex minimizer. The significant fea-
ture of both plots is that the docked pose generally
scores more favorably than the minimized crystal
structure. When rigid ligand docking is compared with
flexible ligand docking, the flexibly docked ligand
conformations almost always have a lower score
(Fig. 8c). These results indicate that increasing the
amount of ligand orientational and conformational
sampling increasingly identifies deeper wells in the
binding energy landscape. When we plotted the flexible
ligand success rate against the minimized crystal score,
there was little correlation, though DOCK was ob-
served to perform better using crystal structures with
scores more negative than –20 DOCK units (Fig. 8d).
This lack of correlation indicates that, while having a
negative interaction energy for the crystal structure
will increase the probability of DOCK finding the
correct binding orientation, this metric is not a good
predictive indicator of DOCKing success.
Database docking using MPI
Substantial speedup is observed for up to about 14
processors for the 500 compound library and 18 pro-
cessors for the 1,000 compound library (Fig. 9). Inter-
estingly, the library with 1,000 copies of neviripine
shows almost perfectly parallel behavior up to 68
processors. We hypothesize that the speedup for the
heterogeneous libraries will continue to approach ideal
as larger libraries with increased numbers of rotatable
bonds are used, but will never be completely linear due
to overhead from input and output and lag resulting
from communication between the nodes.
Discussion
In this paper we have described a new version of the
DOCK program. Our main purpose was to develop
modular code that was straightforward to modify and
c)a)
b) d)
Fig. 8 (a) Successful rigid
ligand docking scores (kcal/
mol) as a function of
minimized crystal structure
ligand scores (kcal/mol),
(b) Successful flexible ligand
docking scores (kcal/mol) as a
function of minimized crystal
structure ligand scores (kcal/
mol), (c) Successful flexible
ligand docking energy scores
(kcal/mol) as a function of
successful rigid ligand
docking energy scores (kcal/
mol), (d) Comparison of the
RMSD between all top
ranked flexible ligand
orientations and the
minimized crystal ligand
orientations to the minimized
crystal interaction energy as
measured by the DOCK score
(kcal/mol)
614 J Comput Aided Mol Des (2006) 20:601–619
123
which showed improved performance over the old
version. By using an object-oriented language for
DOCK 5, we were able to accomplish this goal, and we
demonstrate, here, how routines such as the simplex
minimizer and the clustering algorithm can be added or
replaced without changes in other parts of the pro-
gram. The successful parallelization of the calculation
and the addition of post-processing clustering were
simple but useful modifications to the algorithm, which
encourages further investigations and algorithm
experimentation.
The performance of DOCK 5 on a curated test set of
114 protein–ligand complexes proved to be superior to
DOCK 4, with an over-all success rate of 79% for rigid
ligand docking and 72% for flexible ligand docking,
compared with 72 and 42%, respectively for DOCK 4.
We ascribe the improvements to significant changes in
the flexible search sampling and pruning procedures
and to code corrections. The difference in performance
of DOCK 5 for rigid and flexible docking is relatively
modest (79% vs. 72%) even though the search for
flexible ligands includes both configurational and con-
formational spaces. Using the receptor structure to
prune the conformational search tree is clearly a rea-
sonably efficient procedure. Although, the DOCK 5
code takes longer on average to run a calculation than
DOCK 4, we feel this drawback is balanced by the
improved results and the modularity of DOCK 5. Ef-
forts to increase throughput are underway.
We also wish to stress the importance of having a
high quality test set for evaluation of docking pro-
grams. X-ray crystallography typically provides essen-
tial but incomplete data for the calculations we wish to
carry out. For example, in the majority of cases,
hydrogen positions must be determined. In other cases,
critical water molecules must be placed and some
residues need to be modeled where experimental data
is lacking. The ligand conformations may also contain
significant uncertainties. Finally, we must be aware of
the inherent assumptions underlying the force field
parameters used in the molecular modeling steps. All
of these considerations speak to the need for careful
inspection of test set complexes. Our results demon-
strate this issue: the success rate for reconstitution of
the complex geometries was shown to depend on the
nature of the cofactors, the optimization of hydrogen
placements, and the inclusion of critical waters.
The primary result that emerges from the analysis of
the docking failures is that the current force field re-
quires improvement, particularly in the treatment of
metal-containing cofactors. We also note that binding
conformations and configurations are determined by the
free energy of the system while we are only, at best,
estimating the enthalpy. Finally, we do identify a few
situations in flexible ligand docking where the confor-
mational sampling is insufficient. A test set with ligands
containing more than seven rotatable bonds would,
presumably, show an increase in these sampling failures.
We hypothesize that the key weakness is the pruning
algorithm, which we will explore in future studies.
What are the routes to improvement? An obvious
starting point is the use of more accurate methods for
preparing experimental structures, including tools for
accurate pK
a
prediction and de novo identification of
critical waters. For the docking calculation itself, it
would be helpful to improve VDW and electrostatic
parameters for all atoms heavier than oxygen, partic-
ularly for metal atoms. Ideally, one would directly in-
clude charge polarization and ligation geometry in the
force field. In addition, modifications to the force field
to better approximate the free energy—e.g. general-
ized Born or Poisson Boltzmann implicit solvation
electrostatics with surface area corrections to account
for the hydrophobic effect—would also improve mod-
eling accuracy. The DOCK 5 platform is positioned to
enable future developments and work is underway to
incorporate them into future releases.
Conclusions
In this study, we have evaluated a new version of
DOCK. We have found that it predicts binding
geometries of a structurally diverse test set comparably
Fig. 9 Speedup (calculated as length of time for calculation on a
single processor/length of time for calculation on n processors)
for docking a library of 500 different small molecules (
s), 1,000
different small molecules (M), and 1,000 copies of nevaripine (S)
using flexible ligand docking as a function of the number of
processors in MPI mode. A perfectly parallel calculation (–) is
plotted for comparison
J Comput Aided Mol Des (2006) 20:601–619 615
123
to similar algorithms and better than the previous
version of DOCK. Simultaneously, we have thoroughly
explored the sampling portions of the algorithm and
found that the majority of binding pose prediction
failures is a result of scoring function deficiencies. In
further exploration of these failures, we have deter-
mined that the docking success seems to be a function
of whether there are alternative orientations that score
well—as defined by the scoring function—rather than
the interaction energy of the experimental structure
itself. Finally, we have implemented new functional-
ities and shown that they improve the success rates of
both rigid and flexible ligand docking. In general, we
have a new tool that not only performs well on a typical
test set but is an ideal tool to explore any number of
new algorithms in the context of the molecular docking
problem.
Acknowledgements Gratitude is expressed to Dr. Bentley
Strockbine and Sudipto Mukherjee for computational assistance
with MPI calculations. Demetri Moustakas, Natasja Brooijmans,
P. Therese Lang and Irwin D. Kuntz would like to thank the NIH
grant GM 56531 (Paul Ortiz de Montellano, PI) for support. P.
Therese Lang would also like to thank the Burroughs Welcome
Foundation and the American Foundation for Pharmaceutical
Education for additional support. The authors would like to
thank Scott Brozell, Mathew Jacobson, and Brian Shoichet and
members of his group for helpful conversations.
Appendix 1
Rigid docking parameter optimization
The parameters listed in Appendix 1 control the
sampling of ligand poses within the receptor active site
during rigid ligand docking. The parameters that con-
trol the step sizes for the simplex minimizer
(simplex_trans_step, simplex_rot_step, and sim-
plex_tors_step) were optimized in a previous study and
were held at those values [14, 49]. For the remaining
parameters—the number of orientations (max_orien-
tations) and the number of minimization steps (sim-
plex_final_max_iterations)—a series of rigid ligand
docking experiments were performed to optimize the
DOCK score for the top ranking pose averaged over
the entire test set and the success rate, defined as the
orientation of the top ranking pose being within 2 A
˚
heavy atom RMSD from the crystal ligand. The success
rate and DOCK scores initially improved as the num-
ber of orientations and the amount of minimization
increased and then converged (Fig. 10). We selected
the lowest converged values—1,000 orientations and
1,000 minimization steps—as optimal.
Flexible docking parameter optimization
For the more complex flexible ligand algorithm, the
parameter optimization was performed first on the
anchor docking, and the best parameters were then
used for optimizing the growth. The parameters that
control the sampling in both these steps are listed in
Appendix 2. As for rigid ligand docking, the
parameters that control step sizes for the simplex
minimizer were set to the previously defined optimal
values.
Fig. 10 Optimization of parameters for rigid ligand docking.
Parameters of 50 (h), 100 (s), 1,000 (O), and 10,000 (.)
minimization steps (simplex_final_max_iterations) are examined
as a function of the number of orientations (max_orientations)
Appendix 1 Description of and optimized default values for parameters that affect rigid ligand docking
Parameter name Parameter description Value
max_orientations The number of ligand poses sampled by the rigid orienting algorithm 1,000
simplex_score_converge The score threshold used to determine simplex convergence 0.1
simplex_trans_step The maximum initial translation step size for the simplex minimizer 1.0 A
˚
simplex_rot_step The maximum initial rotational euler angle step size for the simplex minimizer 0.1 radian
simplex_tors_step The maximum initial dihedral angle step size 10
simplex_final_max_iterations The maximum number of simplex iterations 1,000
616 J Comput Aided Mol Des (2006) 20:601–619
123
The first step in the anchor-and-grow algorithm is
ring identification or anchor segmentation. All bonds
within molecular rings are treated as rigid. This clas-
sification scheme is a first-order approximation of
molecular flexibility, since some amount of flexibility
can exist in non-aromatic rings. To treat such phe-
nomena as sugar puckering and chair-boat hexane
conformations, the user needs to supply each ring
conformation as a separate input molecule. If the
molecule does not have a ring, the largest rigid seg-
ment is specified as the anchor. Additional bonds may
be specified as rigid by the user. For simplicity, all runs
in this study used the default of largest anchor only. If
the molecule had multiple anchors of the same size, the
first anchor on the anchor list was used. Once the an-
chor had been identified, the parameters that control
the number of anchor orientations (max_orientations),
the number of anchor minimization steps (sim-
plex_anchor_max_iterations), and the cutoff for the
anchor pruning (num_confs_for_next_growth) were
explored. Because the anchors are substructures of the
ligand, the parameter convergence was monitored as a
function of the RMSD between the anchor orientation
and the corresponding substructure of the crystal li-
gand averaged over all generated orientations before
the pruning function. When the number of anchor
orientations and minimization steps were varied sys-
tematically, the number of minimization steps con-
verged at 500 (Fig. 11a). We expected this optimized
value to be lower than rigid docking because anchors
are typically smaller than the final ligand.
Because the anchor orientations are pruned before
the growth step, we used the optimized number of
minimization steps while exploring the number of
anchor orientations and the pruning cutoff. The
optimal anchor pruning cutoff of 100 was chosen as a
balance between convergence and the length of the
calculation, which remained fixed for the final explo-
ration of the number of orientations. The optimal
number of orientations was selected to be 500 because
the combination of these three variables generated
the highest number of anchors near the crystal
structure (Fig. 11a). Note that if the number of ori-
entations was increased beyond the selected value,
the number of anchors near the crystal structure
dropped dramatically. We hypothesized that this
resulted from a combination of increased sampling
and pruning. The pruning function was designed to
identify a representative orientation from each energy
well that the matching algorithm finds (see Introduc-
tion: DOCK background). As sampling increased, the
ranked orientations began to converge toward the
bottom of the deepest energy wells, sampling less of
the alternative high energy wells. Because the pruning
function is designed to supply the most diverse
ligands, fewer orientations made it through the
pruning step as the sampling is increased. We felt that
this effect was reducing the potential sampling for the
algorithm and plan to explore alternatives in future
studies.
The next step in the anchor-and-grow algorithm is
flexible bond identification. Each flexible bond is
associated with a label defined in an editable file. The
parameter file is identified with the flex_definition_file
parameter. Each label in the file contains a definition
based on the atom types and chemical environment of
the bonded atoms. Typically, bonds with some degree
of double bond character are excluded from minimi-
zation so that planarity is preserved. Each label is also
associated with a set of preferred torsion positions.
The location of each flexible bond is used to partition
the molecule into rigid segments. A segment is the
largest local set of atoms that contains only non-
flexible bonds.
Using the optimal anchor parameters, we varied
number of minimization steps for each layer of growth
(simplex_grow_max_iterations) and the cutoff of
number of conformers for the growth pruning function
(num_confs_for_next_growth). Because the dock run
now creates a complete pose, we return to using a
combination of the score for the top ranking pose
averaged over the entire test set and the success rate to
Appendix 2 Description of and optimized default values for parameters that affect flexible ligand docking
Parameter name Parameter description Value
max_orientations The number of anchor poses sampled by the rigid orienting algorithm 500
num_anchor_orients_for_growth The maximum number of anchor orientations promoted to the conformational search 100
num_confs_for_next_growth The number of partially grown ligand conformers stored at each stage of the flexible growth
procedure
100
simplex_anchor_max_iterations The maximum number of simplex iterations applied to the ligand anchor during anchor
docking
500
simplex_grow_max_iterations The maximum number of simplex iterations applied to the ligand during the flexible growth
procedure
500
J Comput Aided Mol Des (2006) 20:601–619 617
123
monitor convergence. As with rigid ligand docking, the
success rate improves modestly with improved sam-
pling and eventually converges (Fig. 11). However,
although DOCK scores improved as the number of
orientations and the amount of minimization in-
creased, the values do not converge. We once again
attribute this phenomenon to the pruning function.
Therefore, we used the success rate to select the lowest
converged values—500 minimization steps and the
cutoff for the number of conformers for the growth
section as 100—as optimal.
References
1. Kopec KK, Bozyczko-Coyne D, Williams M (2005) Biochem
Pharmacol 69:1133
2. Congreve M, Murray CW, Blundell TL (2005) Drug Dis-
covery Today 10:895
3. Kraljevic S, Stambrook PJ, Pavelic K (2004) EMBO Rep 5:837
4. Schnecke V, Bostrom J (2006) Drug Discovery Today 11:43
5. Hillisch A, Pineda LF, Hilgenfeld R (2004) Drug Discovery
Today 9:659
6. Posner BA (2005) Curr Opin Drug Discovery Dev 8:487
7. Alvarez JC (2004) Curr Opin Chem Biol 8:365
8. Verdonk ML, Cole JC, Hartshorn MJ, Murray CW, Taylor
RD (2003) Proteins 52:609
9. Friesner RA, Banks JL, Murphy RB, Halgren TA, Klicic JJ,
Mainz DT, Repasky MP, Knoll EH, Shelley M, Perry JK,
Shaw DE, Francis P, Shenkin PS (2004) J Med Chem 47:1739
10. Halgren TA, Murphy RB, Friesner RA, Beard HS, Frye LL,
Pollard WT, Banks JL (2004) J Med Chem 47:1750
11. Kramer B, Rarey M, Lengauer T (1999) Proteins 37:228
12. Kitchen DB, Decornez H, Furr JR, Bajorath J (2004) Nat
Rev Drug Discovery 3:935
13. Shoichet BK, Bodian DL, Kuntz ID (1992) J Comput Chem
13:380
14. Ewing TJA, Kuntz ID (1997) J Comput Chem 18:1175
15. Leach AR, Kuntz ID (1992) J Comput Chem 13:730
16. Meng EC, Shoichet BK, Kuntz ID (1992) J Comput Chem
13:505
17. Lischner R (2003) C++ in a nutshell. 1st edn. O’Reilly
Media, Inc, Sebastopol, CA
a)
b)
Fig. 11 Optimization of parameters for flexible ligand docking.
(a) Parameter optimization for anchor sampling portion of
flexible ligand docking. TOP: Parameters of 0 (h), 50 (s), 100
(n), and 500 (O) anchor minimization steps (sim-
plex_anchor_max_iterations) are plotted as a function of the
number of orientations (max_orientations). BOTTOM: Param-
eters of 50 (vertical stripes), 500 (filled), and 5,000 (diagonal
stripes) anchor orientations (max_orientations) are compared
using an anchor pruning cutoff (num_confs_for_next_growth) of
100. (b) Parameter optimization for growth sampling portion of
flexible ligand docking. Growth pruning cutoffs (num_con-
fs_for_next_growth) of 25 (s), 50 (n), 100 (O), and 200 (e)
are plotted as a function of the number of growth minimization
steps (simplex_grow_max_iterations)
618 J Comput Aided Mol Des (2006) 20:601–619
123
18. Berman HM, Westbrook J, Feng Z, Gilliland G, Bhat TN,
Weissig H, Shindyalov IN, Bourne PE (2000) Nucleic Acids
Res 28:235
19. Jones G, Willett P, Glen RC, Leach AR, Taylor R (1997) J
Mol Biol 267:727
20. Pang YP, Perola E, Xu K, Prendergast FG (2001) J Comput
Chem 22:1750
21. Perola E, Walters WP, Charifson PS (2004) Proteins 56:235
22. Nissink JW, Murray C, Hartshorn M, Verdonk ML, Cole JC,
Taylor R (2002) Proteins 49:457
23. Kuhl FS, Crippen GM, Friesen DK (1984) J Comput Chem
5:24
24. Nelder JA, Mead R (1965) Comput J 7:308
25. Gropp W, Lusk E, Doss N, Skjellum A (1996) Parallel
Computing 22:789
26. SYBYL, Tripos, Inc., St. Louis, Missouri, 63144
27. Case DA, Darden TA, Cheatham III, TE, Simmerling CL,
Wang J, Duke RE, Luo R, Merz KM, Wang B, Pearlman
DA, Crowley M, Brozell S, Tsui V, Gohlke H, Mongan J,
Hornak V, Cui G, Beroza P, Schafmeister C, Caldwell JW,
Ross WS, Kollman PA (2004) AMBER 8, University of
California, San Francisco
28. Jakalian A, Bush BL, Jack DB, Bayly CI (2000) J Comput
Chem 21:132
29. Hann MM, Oprea TI (2004) Curr Opin Chem Biol 8:255
30. Oprea TI (2002) J Comput-Aided Mol Des 16:325
31. Oprea TI, Davis AM, Teague SJ, Leeson PD (2001) J Chem
Inf Model 41:1308
32. Brooijmans N (2003) Theoretical studies of molecular rec-
ognition, Graduate Department of Chemistry and Chemical
Biology, University of California, San Francisco, San Fran-
cisco, CA
33. Purcell WP, Singer JA (1967) J Chem Eng Data 12:235
34. Gasteiger J, Marsili M (1980) Tetrahedron 36:3219
35. Aqvist J, Warshel A (1990) J Am Chem Soc 112:2860
36. Merz KM, Murcko MA, Kollman PA (1991) J Am Chem Soc
113:4484
37. Cornell WD, Cieplak P, Bayly CI, Gould IR, Merz KM,
Ferguson DM, Spellmeyer DC, Fox T, Caldwell JW, Koll-
man PA (1995) J Am Chem Soc 117:5179
38. Richards FM (1977) Ann Rev Biophys Bioeng 6:151
39. DesJarlais RL, Sheridan RP, Seibel GL, Dixon JS, Kuntz ID,
Venkataraghavan R (1988) J Med Chem 31:722
40. Kuntz ID, Blaney JM, Oatley SJ, Langridge R, Ferrin TE
(1982) J Mol Biol 161:269
41. Pettersen EF, Goddard TD, Huang CC, Couch GS, Green-
blatt DM, Meng EC, Ferrin TE (2004) J Comput Chem
25:1605
42. Meng EC, Lewis RA (1991) J Comput Chem 12:891
43. Mills JEJ, Dean PM (1996) J Comput-Aided Mol Des 10:607
44. Irwin JJ, Shoichet BK (2005) J Chem Inf Model 45:177
45. The results for the FlexX test set are available at http://
www.biosolveit.de/FlexX/
46. The results for the GOLD test set are available at http://
www.ccdc.cam.ac.uk/products/life_sciences/validate/
gold_validation/value.html
47. Verkhivker GM, Bouzida D, Gehlhaar DK, Rejto PA, Ar-
thurs S, Colson AB, Freer ST, Larson V, Luty BA, Marrone
T, Rose PW (2000) J Comput-Aided Mol Des 14:731
48. Kuntz ID, Agard DA (2003) Adv Protein Chem 66:1
49. Gschwend DA, Kuntz ID (1996) J Comput-Aided Mol Des
10:123
J Comput Aided Mol Des (2006) 20:601–619 619
123
