R E V I E W S
NATURE REVIEWS | DRUG DISCOVERY
VOLUME 3 | NOVEMBER 2004 | 935
The number of proteins with a known three-dimen-
sional structure is increasing rapidly,and structures pro-
duced by structural genomics initiatives are beginning to
become publicly available1,2.The increase in the number
of structural targets is in part due to improvements in
techniques for structure determination,such as high-
throughput X-ray crystallography3. With large-scale
structure-determination projects driven by genomics
consortia, many current target proteins have been
selected for their therapeuticpotential.
Computational methodologies have become a crucial
component of many drug discovery programmes,
from hit identification to lead optimization and
beyond4–6,and approaches such as ligand-4or structure-
based virtual screening7techniques are widely used in
many discovery efforts. One key methodology —
docking of small molecules to protein binding sites —
was pioneered during the early 1980s8,and remains a
highly active area of research7.When only the structure
of a target and its active or binding site is available,
high-throughput docking is primarily used as a hit-
identification tool.However,similar calculations are
often also used later on during lead optimization,
when modifications to known active structures can
quickly be tested in computer models before compound
synthesis.Furthermore,docking can also contribute to
the analysis of drug metabolism using structures such
as cytochrome P450 isoforms9,10.
Here,we review basic concepts and specific features of
small-molecule–protein docking methods and several
selected applications,with particular emphasis on hit
identification and lead optimization,but do not specifi-
cally review protein–protein docking,which is less rele-
vant for small-molecule drug discovery.We attempt to
distinguish between the problems of docking com-
pounds into target sites and ofscoring docked conforma-
tions,because the available data indicate that numerous
robust and accurate docking algorithms are available,
whereas imperfections ofscoring functions continue to
be a major limiting factor.
An introduction to docking
The docking process involves the prediction of ligand
conformation and orientation (or posing) within a
targeted binding site (BOX 1).In general,there are two
aims of docking studies:accurate structural modelling
and correct prediction of activity.However,the identifi-
cation of molecular features that are responsible for
specific biological recognition, or the prediction of
compound modifications that improve potency,are
DOCKING AND SCORING IN VIRTUAL
SCREENING FOR DRUG DISCOVERY:
METHODS AND APPLICATIONS
Douglas B.Kitchen*,Hélène Decornez*,John R.Furr* and Jürgen Bajorath‡,§
Abstract | Computational approaches that ‘dock’ small molecules into the structures of
macromolecular targets and ‘score’ their potential complementarity to binding sites are widely
used in hit identification and lead optimization. Indeed, there are now a number of drugs whose
development was heavily influenced by or based on structure-based design and screening
strategies, such as HIV protease inhibitors. Nevertheless, there remain significant challenges
in the application of these approaches, in particular in relation to current scoring schemes.
Here, we review key concepts and specific features of small-molecule–protein docking
methods, highlight selected applications and discuss recent advances that aim to address
the acknowledged limitations of established approaches.
21 Corporate Circle,Albany,
‡AMRI Bothell Research
18804 North Creek Parkway,
Correspondence to J.B.
The process of determining
whether a given conformation
and orientation of a ligand fits
the active site.This is usually a
fuzzy procedure that returns
many alternative results.
Both posing and ranking involve
scoring.The pose score is often
a rough measure of the fit of a
ligand into the active site.The
rank score is generally more
complex and might attempt to
estimate binding energies.
A more advanced process than
pose scoring that typically takes
several results from an initial
scoring phase and re-evaluates
them.This process usually
attempts to estimate the free
energy of binding as accurately
as possible.Although the posing
phase might use simple energy
calculations (electrostatic and
van der Waals),ranking
procedures typically involve more
elaborate calculations (perhaps
including properties such as
entropy or explicit solvation).
complex issues that are often difficult to understand and
— even more so — to simulate on a computer.
In view of these challenges, docking is generally
devised as a multi-step process in which each step intro-
duces one or more additional degrees of complexity11.
The process begins with the application of docking
algorithms that POSEsmall molecules in the active site.
This in itself is challenging,as even relatively simple
organic molecules can contain many conformational
degrees of freedom.Sampling these degrees of freedom
must be performed with sufficient accuracy to identify
the conformation that best matches the receptor struc-
ture,and must be fast enough to permit the evaluation
of thousands of compounds in a given docking run.
Algorithms are complemented by SCORING FUNCTIONS that
are designed to predict the biological activity through
the evaluation of interactions between compounds and
potential targets. Early scoring functions evaluated
compound fits on the basis of calculations of approxi-
mate shape and electrostatic complementarities.
Relatively simple scoring functions continue to be
heavily used,at least during the early stages of docking
simulations.Pre-selected conformers are often further
evaluated using more complex scoring schemes with
more detailed treatment of electrostatic and van der
Waals interactions,and inclusion of at least some solva-
tion or entropic effects7.It should also be noted that
ligand-binding events are driven by a combination of
enthalpic and entropic effects,and that either entropy
or enthalpy can dominate specific interactions.This
often presents a conceptual problem for contemporary
scoring functions (discussed below),because most of
∆G = –RT1nKA
KA = Ki–1 =
[E]aq + [I]aq
[E + I]aq
936 | NOVEMBER 2004 | VOLUME 3
R E V I E W S
them are much more focused on capturing energetic
than entropic effects.
In addition to problems associated with scoring of
compound conformations,other complications exist that
make it challenging to accurately predict binding confor-
mations and compound activity.These include,among
others, limited resolution of crystallographic targets,
inherent flexibility,induced fit or other conformational
changes that occur on binding,and the participation of
water molecules in protein–ligand interactions.Without
doubt,the docking process is scientifically complex.
Molecular representations for docking
To evaluate various docking methods,it is important to
consider how the protein and ligand are represented.
There are three basic representations of the receptor:
atomic,surface and grid12.Among these,atomic represen-
tation is generally only used in conjunction with a poten-
tial energy function13and often only during final RANKING
procedures (because ofthe computational complexity of
evaluating pair-wise atomic interactions).
Surface-based docking programs are typically,but
not exclusively,used in protein–protein docking14,15.
Connolly’s early work on molecular surface representa-
tions is mainly responsible for spawning much of the
research in this area16,17.These methods attempt to align
points on surfaces by minimizing the angle between the
surfaces of opposing molecules18.Therefore,a rigid
body approximation is still the standard for many
protein–protein docking techniques.
The use of potential energy grids was pioneered by
Goodford19,and various docking programs use such
grid representations for energy calculations.The basic
idea is to store information about the receptor’s ener-
getic contributions on grid points so that it only needs
to be read during ligand scoring. In the most basic
form,grid points store two types of potentials:electro-
static and van der Waals (BOX 2). FIGURE 1shows a repre-
sentative grid for capturing electrostatic potentials,and
FIG.2 illustrates the electrostatic potential of a bound
inhibitor mapped on its molecular surface.
Search methods and molecular flexibility
This section focuses on algorithms used to treat ligand
flexibility and, to some extent, protein flexibility.
Treatment ofligand flexibility can be divided into three
basic categories11:systematic methods (incremental con-
struction,conformational search,databases);random or
stochastic methods (Monte Carlo,genetic algorithms,
tabu search); and simulation methods (molecular
dynamics, energy minimization). A summary of the
search approaches implemented in widely used docking
programs is presented in BOX 3.
Systematic search.These algorithms try to explore all the
degrees offreedom in a molecule,but ultimately face the
problem ofcombinatorial explosion20(BOX 4).Therefore,
ligands are often incrementally grown into active sites.
A stepwise or incremental search can be accomplished in
different ways — for example,by docking various molec-
ular fragments into the active-site region and linking
Box 1 | Theoretical aspects of docking
For an enzyme and inhibitor,docking aims at correct prediction ofthe structure ofthe
complex [E+I] = [EI] under equilibrium conditions (see figure and equation 1).
The figure illustrates the binding ofinhibitor Dmp323 to HIV protease and is based on
solution structures (PDB code:1BVE).Multiple structures ofenzyme–inhibitor
complexes revealed only limited structural variations.
The free energy ofbinding (∆G) is related to binding affinity by equations 2 and 3:
Prediction ofthe correct structure (posing) ofthe [E+I] complex does not require
information about KA.However,prediction ofbiological activity (ranking) requires this
information;scoring terms can therefore be divided in the following fashion.When
considering the term [EI],the following factors are important:steric,electrostatic,
hydrogen bonding,inhibitor strain (ifflexible) and enzyme strain.When considering the
equilibrium shown in equation 1,the following factors are also important:desolvation,
rotational entropy and translational entropy.
NATURE REVIEWS | DRUG DISCOVERY
VOLUME 3 | NOVEMBER 2004 | 937
R E V I E W S
stage,energy minimization is performed after each
Another method of systematic search is the use of
libraries ofpre-generated conformations.Library confor-
mations are typically only calculated once and the search
problem is therefore reduced to a rigid body docking
procedure.For example,FLOG29generates database con-
formations on the basis of distance geometry. Once
acceptable conformations have been generated,the algo-
rithm explores them in a manner similar to DOCK11,29.
Random search.These algorithms (often called stochastic
methods) operate by making random changes to either a
single ligand or a population ofligands.A newly obtained
ligand is evaluated on the basis of a pre-defined proba-
bility function.Two popular random approaches are
Monte Carlo and genetic algorithms (BOX 5).Alternative
implementations of Monte Carlo search have been
reported30,31,including a popular form in AutoDock30.By
contrast,several other programs (including DOCK and
GOLD) have implemented genetic algorithms32–34.
The basic idea of a tabu search algorithm is to take
into consideration already explored areas of conforma-
tional space35,36.To determine whether a molecular con-
formation is accepted or not, the root mean square
deviation is calculated between current molecular
coordinates and every molecule’s previously recorded
conformation.For example,PRO_LEADS makes use
of a tabu search algorithm35.
Simulation methods.Molecular dynamics is currently
the most popular simulation approach.However,molec-
ular dynamics simulations are often unable to cross
high-energy barriers within feasible simulation time
periods,and therefore might only accommodate ligands
in local minima of the energy surface11.Therefore,an
attempt is often made to simulate different parts of a
protein–ligand system at different temperatures37.
Another strategy for addressing the local minima prob-
lem is starting molecular dynamics calculations from
different ligand positions. In contrast to molecular
dynamics,energy minimization methods are rarely used
as stand-alone search techniques,as only local energy
minima can be reached,but often complement other
search methods,including Monte Carlo38.DOCK per-
forms a minimization step after each fragment addition,
followed by a final minimization before scoring.
Protein flexibility.The treatment ofprotein flexibility is
less advanced than that of ligand flexibility,but various
approaches have been applied to flexibly model at least
part of the target39,including molecular dynamics and
Monte Carlo calculations31–33,rotamer libaries40,41and
protein ensemble grids42.The idea behind using amino-
acid side-chain rotamer libraries is to model protein
conformational space on the basis of a limited number
of experimentally observed and preferred side-chain
conformations40.To reduce the number ofdiscrete pro-
tein conformations arising from combinations of
rotamers,a dead-end elimination algorithm is often
used41.This algorithm recursively removes side-chain
them covalently (which is most popular as a de novo
ligand-design strategy) or, alternatively, by dividing
docked ligands into rigid (core fragment) and flexible
parts (side chains).In the latter case,once the rigid cores
have been defined,they are docked into the active site.
Next, flexible regions are added in an incremental
fashion21–23. For example,DOCK 4.024poses the core
fragment by steric complementarity,and flexible side
chains are grown one bond at a time by systematically
exploring each bond’s POSE SPACE.A pruning algorithm is
applied to remove unfavourable conformations early
on,thereby reducing the complexity of the problem24,25.
FlexX differs from DOCK in that the placement of the
rigid core fragment is based on interaction geometries
between fragments and receptor groups22,26.Interacting
groups are primarily hydrogen-bond donors and accep-
tors,as well as hydrophobic groups.FlexX further differs
from DOCK in that it uses a pose-clustering algorithm
to classify the docked poses22,27.
The Hammerhead algorithm28,in common with other
incremental search algorithms,also divides ligands into
fragments.However,Hammerhead docks each fragment
and then rebuilds the ligand from fragments that have
acceptable initial scores.During the fragment-growing
All degrees of freedom involved
in the process of placing one
molecule relative to another.
For example,for two rigid
molecules the pose space simply
consists of relative orientations.
When one of the molecules,the
ligand,is allowed to be flexible,
the pose space comprises both
the conformational space of the
ligand and orientational space
of ligand and receptor.
Ecoul(r) = (1)
i = 1
j = 1
EvdW(r) = –
j = 1
i = 1
Van der Waals energy
Box 2 | Standard potential energy functions
The electrostatic potential energy is represented as a pair-
wise summation ofCoulombic interactions,as described
in equation 1:
In equation 1,Nis the number ofatoms in molecules A
and B,respectively,and qthe charge on each atom.
The van der Waals potential energy for the general
treatment of non-bonded interactions is often modelled
by a Lennard–Jones 12–6 function,as shown in
In equation 2,εis the well depth of the potential and σis
the collision diameter ofthe respective atoms i and j.
The figure shows a representation ofthe Lennard–Jones
12–6 function.The exp(12) term of the equation is
responsible for small-distance repulsion,whereas the
exp(6) provides an attractive term which approaches
zero as the distance between the two atoms increases.
938 | NOVEMBER 2004 | VOLUME 3
R E V I E W S
energy (such as steric strain induced by binding).FIGURE 3
illustrates force-field modelling of non-bonded inter-
actions involved in molecular recognition.Most force-
field scoring functions only consider a single protein
conformation,which makes it possible to omit the calcu-
lation ofinternal protein energy,which greatly simplifies
scoring.Various force-field scoring functions are based
on different force field parameter sets.For example,G-
Score26is based on the Tripos force field26and AutoDock45
on the AMBER force field46. However, functional
forms are usually similar,as shown in supplementary
information S1 (table).
Interactions between ligand and receptor are most
often described by using van der Waals and electro-
static energy terms.The van der Waals energy term is
given by a Lennard–Jones potential function (BOX 2).
The parameters of the Lennard–Jones potential vary
depending on the desired ‘hardness’of the potential.
Higher terms,such as a 12–6 Lennard–Jones potential of
D-Score26,result in increasingly repulsive potentials
and will be less forgiving of close contacts between
receptor and ligand atoms.Accordingly,lower terms,
such as the 8–4 Lennard–Jones potential of G-score26,
make the potential softer. Electrostatic terms are
accounted for by a Coulombic formulation with a dis-
tance-dependent dielectric function that lessens the
contribution from charge–charge interactions.The
functional form of the internal ligand energy is typi-
cally very similar to the protein–ligand interaction
energy,and also includes van der Waals contributions
and/or electrostatic terms.
Standard force-field scoring functions have major
limitations,because they were originally formulated to
model enthalpic gas-phase contributions to structure
and energetics, and do not include solvation and
entropic terms.Force-field-based scoring is further com-
plicated by the fact that it generally requires the introduc-
tion ofcut-offdistances for the treatment ofnon-bonded
interactions,which are more or less arbitrarily chosen
conformations that do not contribute to a minimum-
energy structure.Another method of treating protein
flexibility is to use ensembles of protein conformations
(rather than a single one) as the target for docking42and
to map these ensembles on a grid representation.One
approach generates an average potential energy grid of
the ensemble,as first implemented in DOCK42;another
maps various receptor potentials to each grid point and
subsequently scores ligand conformations against each
set ofreceptor potentials19.
The evaluation and ranking ofpredicted ligand confor-
mations is a crucial aspect of structure-based virtual
screening.Even when binding conformations are cor-
rectly predicted, the calculations ultimately do not
succeed if they do not differentiate correct poses from
incorrect ones,and if‘true’ligands cannot be identified.
So,the design ofreliable scoring functions and schemes is
of fundamental importance. Free-energy simulation
techniques have been developed for quantitative model-
ling ofprotein–ligand interactions and the prediction of
binding affinity43,44.However,these expensive calculations
remain impractical for the evaluation oflarge numbers of
protein–ligand complexes and are not always accurate.
Scoring functions implemented in docking programs
make various assumptions and simplifications in the
evaluation of modelled complexes and do not fully
account for a number ofphysical phenomena that deter-
mine molecular recognition — for example,entropic
effects.Essentially,three types or classes ofscoring func-
tions are currently applied: FORCE-FIELD-based,empirical
and knowledge-based scoring functions.BOX 6summa-
rizes a number ofcurrently used scoring functions,details
ofwhich will be discussed in the following section.
Force-field-based scoring.Molecular mechanics force
fields usually quantify the sum of two energies, the
receptor–ligand interaction energy and internal ligand
A function expressing the
energy of a system as a sum of
diverse molecular mechanics
(or other) terms.
Figure 1 | Grid representations. a | Shown is a surface plot of a grid capturing the electrostatic potential of HIV protease (PDB
code: 1BVE) around its active site (with bound inhibitor Dmp323). Red and blue indicate areas of negative and positive
electrostatic potential, respectively. b | Shows a ‘cut-away’ electrostatic potential grid of the enzyme around the bound inhibitor
(not included in the calculation).
NATURE REVIEWS | DRUG DISCOVERY
VOLUME 3 | NOVEMBER 2004 | 939
R E V I E W S
which is used for peptide docking,explicitly takes ligand
desolvation into account,and desolvation energies are
calculated using a continuum electrostatic model52,53.
However,terms currently used to approximate entropy
or desolvation energy provide only incomplete descrip-
tions of these effects on protein–ligand binding.
Knowledge-based scoring functions.In essence,know-
ledge-based scoring functions are designed to reproduce
experimental structures rather than binding energies.In
knowledge-based functions,as shown in supplementary
information S3(table),protein–ligand complexes are
modelled using relatively simple atomic interaction-pair
potentials. A number of atom-type interactions are
defined depending on their molecular environment.So,
in common with empirical methods,knowledge-based
scoring functions attempt to implicitly capture binding
effects that are difficult to model explicitly. Popular
implementations ofsuch functions include POTENTIAL OF
MEAN FORCE (PMF)54–56and DrugScore57, which also
includes solvent-accessibility corrections to pair-wise
potentials.SMoG58is another scoring function belong-
ing to this class that utilizes pair-wise atom potentials to
evaluate protein–ligand interactions.A major attraction
of many knowledge-based scoring functions is their
computational simplicity, which permits efficient
screening oflarge compound databases.A disadvantage
is that their derivation is essentially based on informa-
tion implicitly encoded in limited sets ofprotein–ligand
Consensus scoring.Given the imperfections of current
scoring functions,a recent trend in this field has been the
introduction ofconsensus scoring schemes59.Consensus
scoring combines information from different scores to
balance errors in single scores and improve the proba-
bility of identifying ‘true’ ligands. An exemplary
implementation of consensus scoring is X-CSCORE60,
which combines GOLD-like,DOCK-like,ChemScore,
PMF and FlexX scoring functions.However,the potential
and complicate the accurate treatment of long-range
effects involved in binding.
Recent extensions of force-field-based scoring
functions include a TORSIONAL ENTROPY term for ligands in
G-Score and the inclusion of explicit protein–ligand
hydrogen-bonding terms in Gold47and AutoDock45.The
latter terms are thought to increase the potential of
specific molecular recognition. Hydrogen-bonding
terms are often designed in noticeably different ways.For
example,G-Score includes different hydrogen-bonding
terms depending on the nature and geometry of the
interaction.By contrast,AutoDock represents all of
the hydrogen bonds by a 12–10 Lennard–Jones
potential with a directional component,as shown in
supplementary information S1 (table).
Empirical scoring functions.These scoring functions are
fit to reproduce experimental data,such as binding ener-
gies and/or conformations,as a sum of several parame-
terized functions, as first proposed by Böhm48. The
design ofempirical scoring functions is based on the idea
that binding energies can be approximated by a sum of
individual uncorrelated terms.The coefficients of the
various terms are obtained from REGRESSION ANALYSIS using
experimentally determined binding energies and,poten-
tially, X-ray structural information. Representative
examples of empirical scoring functions are given in
supplementary information S2(table).The functional
forms are often simpler than force-field scoring func-
tions, although many of the individual contributing
terms have counterparts in the force-field molecular
mechanics terms.The appeal of empirical functions is
that their terms are often simple to evaluate,but they are
based on approximations similar to force-field functions.
A disadvantage ofthese methods is their dependence on
the molecular data sets used to perform regression
analyses and fitting.This often yields different weighting
factors for the various terms.As a consequence,terms
from differently fitted scoring functions cannot easily be
recombined into a new scoring function.
In empirical scoring functions,terms accounting
for non-bonded interactions can be implemented in
rather different ways.For example,in the early LUDI
formulation48,the hydrogen-bonding term is separated
into neutral hydrogen bonds and ionic hydrogen bonds,
whereas ChemScore49does not differentiate between
different types of hydrogen bonds. Furthermore, the
LUDI function calculates hydrophobic contributions on
the basis of a representation of molecular surface area,
whereas ChemScore evaluates contacts between
hydrophobic atom pairs.F-Score adds an additional term
to account for aromatic interactions50.
Empirical scoring functions can include non-
enthalpic contributions such as the so-called rotor term,
which approximates entropy penalties on binding from a
weighted sum of the number of rotatable bonds in lig-
ands.ChemScore implements ligand rotational entropy
in a more complicated form that describes the molecular
environment surrounding each rotatable bond.More
complex functions begin to address solvation and desol-
vation effects.For example,the Fresno scoring function51,
Entropy associated with a
rotatable bond in a molecule.
Immobilization of a rotatable
bond on binding leads to loss
of its torsional (or rotational)
Determination of parameter
values for a chosen (linear or
nonlinear) function to best fit
a set of observations.
POTENTIAL OF MEAN FORCE
(PMF).In the context of docking
and scoring,PMFs are derived
from statistical analysis of
distributions and frequencies of
specific atom-pair interactions
in a large collection of
Interaction potentials between
each atom pair in two molecules
(for example,ligand and
protein) approximate the free
energy of each pair-wise
interaction as a function of
Figure 2 | Electrostatic potential of a bound inhibitor.
Inhibitor Dmp323 is shown in complex with HIV protease (PDB
code: 1BVE). The electrostatic potential of the symmetrical
inhibitor in its binding conformation was mapped on its
calculated molecular surface. Residues Ile50 and Asp25 from
each monomer in HIV protease stabilize inhibitor binding.
940 | NOVEMBER 2004 | VOLUME 3
R E V I E W S
improved the results over AutoDock.These findings also
illustrated the difficulty in finding correct poses for
highly flexible ligands.
Charifson et al.59showed that most conventional
scoring functions were able to place ~50% of active
(<100 nM) compounds within the top 1,000 of the
score lists but that less active (~1 µM) compounds were
much more difficult to identify.Chemscore,the DOCK
energy and PLPscores performed well for all three bind-
ing sites that were analysed.In this study,a consensus
scoring scheme was applied that combined results from
each scoring function and was found to further improve
prediction accuracy. In a similar study,Wang et al.66
analysed more than 100 protein–ligand complexes.The
best scoring conformations from PLP,F-Score,LigScore
and DrugScore were within 2 Å in 70–80% of the test
cases.In this study,the quality of scoring schemes was
found to be more or less independent of the type of
active site.However,scoring functions that relied purely
on force-field or knowledge-based potentials consis-
tently performed worse in identifying correct poses than
functions that used additional terms accounting for
The FRED program uses a Gaussian function for
docking67to generate a smooth and easily searchable
energy surface,and allows a wide latitude for errors in
positioning protein atoms.Pre-generation of multiple
ligand conformers is a suitable technique when this ‘soft’
and error-tolerant function is used,but the method has
limited accuracy in ranking ligands correctly (and
therefore requires more detailed follow-up scoring).In
a comparison of FRED and Glide scoring schemes68,
kinase, gyrase B, thrombin, gelatinase-A and neur-
aminidase binding sites were used as test cases. As
expected,‘hard’functions such as the one in Glide per-
formed overall better than the softer scoring function in
FRED.However,FRED was found to produce accurate
results for lipophilic binding sites, especially when
hydrophobic effects outweighed electrostatic and hydro-
gen-bonding interactions.Scoring functions therefore
respond differently to specific features in binding sites.
Posing versus scoring.Are calculation errors more asso-
ciated with predicting binding conformations or with
scoring them? Ligand flexibility has a greater effect on
predicting structures correctly than size or polarity69,
which clearly relates to posing.However,the ability to dis-
criminate docked conformations of‘fantasy’ligands from
‘true’ones depends crucially on scoring.Clearly,it is often
difficult to distinguish between inadequate conforma-
tional searching and flawed scoring,and relatively few
studies have been designed to dissect these effects.In a
parallel application ofvarious programs70,DOCK,FlexX
and GOLD displayed a clear tendency for ligands to score
better when using X-ray conformationsthan any incor-
rectly modelled conformation, whereas CDOCKER
often scored correct structures worse,thereby indicating
shortcomings of its scoring function. In terms of
structural accuracy,GOLD and Glide produced over-
all satisfactory results for a set of 69 complexes. In a
value of consensus scoring might be limited,if terms in
different scoring functions are significantly correlated,
which could amplify calculation errors,rather than
Evaluating scoring schemes.Perez et al.61compared force-
field scoring with a combined PMF knowledge-based
function.In this study,force-field terms generally per-
formed better than PMF,but steric contributions sig-
nificantly outweighed electrostatic terms.Correct ligand
poses were found in nearly 80% ofthe cases studied when
the force-field function was used,but the success rate
dropped to 56% when cross-docking experiments were
performed.In an effort to better understand incorrectly
calculated effects, LINEAR DISCRIMINANT ANALYSISwas applied;
the results indicated that better adjusted molecular vol-
ume and hydrogen bonding terms were likely to further
improve force-field scoring.In the PMF function,volume
and attractive dispersion effects were not considered.
Good et al.62compared several docking protocols.
When PHARMACOPHOREconstraints and conformational
flexibility were taken into account,simple contact scores
outperformed force-field treatment (even when over-
estimated electrostatic effects were ignored). More
complex functions implemented in Prometheus63and
GOLD64performed better in well-defined active sites,but
simpler schemes such as the DOCK contact energy per-
formed best in more complex and less well-understood
Sotriffer et al.65compared DrugScore (knowledge-
based) with AutoDock (force-field-based) and found
that these two scoring methods had very similar abilities
in predicting the correct binding modes for 158 com-
plexes. For rigid molecules, 90–100% of cases were
predicted correctly.However,as the number ofrotatable
bonds (and molecular flexibility) increased,the success
rate dropped to 44–80%,and DrugScore only marginally
Mathematical analysis based on
two classes of data and two
independent variables (a,b) that
attempts to find a line that best
separates the data.This line is
orthogonal to the discriminant
function that is a linear
combination of the original
variables,in this case:F = caa+
The spatial arrangement ofatoms
or groups in a molecule known
or predicted to be responsible for
specific biological activity.
Box 3 | Flexible ligand-search methods
NATURE REVIEWS | DRUG DISCOVERY
VOLUME 3 | NOVEMBER 2004 | 941
R E V I E W S
across many different protein families,regardless oftheir
complexity and sophistication.Binding sites and recogni-
tion processes have unique features that ultimately render
protein–ligand interactions specific (FIG.4),which in turn
might often require ‘tuning’ofscoring schemes on a case-
by-case basis.A good example of re-evaluating scoring
schemes and adjusting to differences between active sites
is provided by a comparison ofknowledge-based scoring
of various structural classes75.When plotting docking
scores versus the logarithm ofbinding constants,it was
observed that slopes and intercepts significantly differed
for structurally distinct active sites.
Structure-based virtual screening
General caveats. Regardless of the functions applied,
scores are known to scale poorly with molecular mass
and the number of rotatable bonds in compounds76.
Large molecules can form many hypothetical interac-
tions in binding sites and therefore have the tendency to
generate better scores than smaller compounds.On the
other hand,the entropy penalty for immobilization of
rotatable bonds, which is frequently not taken into
account, scales with the number of such bonds. As a
result,ifentropy penalties are included,flexible molecules
tend to score lower than more rigid ones.Furthermore,
the internal strain energy ofa molecular pose is generally
approximated using a single unbound conformation of
the ligand as a reference,which has significant limitations
in estimating entropy and enthalpy losses on binding.
These limitations generally add to imperfections of
scoring functions and make it more difficult to accu-
rately rank test molecules on the basis of computed
The general nature and preparation of active sites
also affects the quality of ligand positions and scores.
Hydrophobic binding sites such as are found in, for
example,HIV protease are likely to be more promising
targets than sites that are more hydrophilic,or binding
events involving distinct electrostatic interactions as seen,
for example,in metallo-enzymes.This is mainly due to
the fact that binding to hydrophobic sites can be well
approximated by a calculation ofshape complementarity
between ligand and receptor,for which robust method-
ologies have existed since the early days of docking8,77.
The calculation of shape complementarity implicitly
takes hydrophobic effects into account;however,a large
(and sometimes the largest) contribution to the
hydrophobic effect comes from desolvation of hydro-
phobic ligands (such as in HIV protease),which is not
adequately accounted for in docking scores and can be
significantly underestimated relative to other scoring
terms in some active sites.Furthermore,precise model-
ling and scoring ofelectrostatic interactions continues to
be a major challenge for contemporary scoring functions.
As mentioned above,simple Coulombic models are still
applied for these purposes in a number ofcases and have
the tendency to grossly overestimate charge–charge
interactions or create artificial ones.
In addition,the placement of water molecules that
are either structurally important or directly involved in
binding interactions,and assumed rigidity ofside-chain
recent re-parameterization of GOLD, two scoring
schemes, GOLDSCORE and ChemScore, were com-
pared and applied in a consensus scoring appproach47.
Both functions yielded similar scoring accuracy
(65–85%). Accurate prediction of relative binding
affinities depended on finding correct binding confor-
mations.As an alternative to consensus scoring,genetic-
algorithm-based scoring terms were introduced to better
distinguish correct ligands from ‘noise’compounds71.On
the basis of these investigations,the conclusion can be
drawn that accurate modelling of binding conforma-
tions is necessary but insufficient for correct ligand
scoring and ranking.Given the success rates of various
efforts at structure prediction, as discussed above, it
seems that imperfections in scoring functions continue
to be a major limiting factor.
Improving scoring functions. How can one further
improve the quality of scoring functions? As indicated
earlier,a current trend in the field is to focus on the inclu-
sion of various solvation52,72,73and rotational entropy69
contributions.Scoring functions accounting for such
contributions are more accurate than,for example,stan-
dard force-field functions,but are also computationally
expensive,which challenges high-throughput docking.
More importantly,however,the often-made observation
that alternative scoring functions perform rather differ-
ently on multiple targets — for example,the GOLD
validation set74— suggests that it might be difficult to
develop scoring functions that perform equally well
NConformations = (1)
i = 1
j = 1
Flexible side chains
Box 4 | The problem of combinatorial explosion
For systematic conformational search,the number of
possible molecular conformations is represented by
In equation 1,Nis the number ofrotatable bonds and θi,j
is the size ofthe incremental rotational angle jfor bond i.
To avoid exhaustive search calculations,many
conformational search algorithms use an incremental
construction approach to grow a ligand within an active
site that consists ofthree basic steps:
• Core fragment selection.
• Core fragment placement.
• Incremental ligand construction.
During the first step,the ligand is divided into a rigid
core fragment and flexible side chains.Subsequently,
these side chains are further divided at each new rotatable
bond,as shown in the figure.
During the second and third steps,core fragments are
placed and side chains are incrementally attached with
rotational degrees of freedom sampled.As discussed in
the text,popular algorithms often differ in the way that
these processes are carried out.
942 | NOVEMBER 2004 | VOLUME 3
R E V I E W S
why structure-based virtual screening actually ‘works’.
The principal reasons are that computational screening is
an enrichment process;that accurately calculated energies
and scores are not necessarily required for meaningful
compound selection; and that appropriate selection
strategies compensate for some methodological short-
comings.For example,in a typical docking study,a large
compound database will probably be reduced to a short-
list ofpreferred candidates,perhaps ~100 or so.To enrich
this selection with compounds that have a high proba-
bility ofbeing active,de-selection ofinappropriate com-
pounds (which most ofthe database compounds are) is
as important as finding the most promising candidates.
Importantly,de-selection ofinappropriate compounds is
more easily achieved than selection within the accuracy
limitations ofthe calculations.Also,some binding events,
such as those dominated by shape complementarity,can
be treated well given the approximations of posing and
scoring.Furthermore,as long as active compounds are
found in the shortlist,their relative ranking becomes less
important.Simply put,an active compound within the
top-five scoring compounds will be as good as one within
the top 50,as long as these compounds are tested,which
further compensates for limitations of scoring.In addi-
tion,it is also a rather common practice to subject a
reasonably small number of pre-selected candidates (for
example, 100–500) to visual inspection, which adds
another dimension to the selection process (that is,
chemical intuition,knowledge and experience)81.So,
although virtual screening inevitably produces false-
positives and -negatives,rationalizing the analysis as an
enrichment process helps to explain its successes.
Even very fast docking and scoring methods typically
require several to tens of seconds per compound for a
fully flexible search and therefore become prohibitive in
the presence of millions of database compounds.As a
result,complex posing and scoring schemes are often
carried out only after the source database has been signifi-
cantly reduced in size by the application ofde-selection or
Structures of target sites. The choice and preparation
of the structural model of a targeted binding site are
important variables.Experimentally determined (X-ray
or nuclear magnetic resonance) structures are generally
preferred.However,as the number of proteins of phar-
maceutical interest has grown faster than the number
whose structures have been determined, homology
modelling has risen in popularity.A recent study com-
pared the quality of docking results when either crystal
structures of HOLO-or APO-ENZYMESor homology models
were used as templates82.Perhaps surprisingly,homol-
ogy models yielded enrichments factors of ten or better
in eight of ten test cases studied,and apo-enzymes and
homology modelled structures performed comparably
well.However,by far the best performance was observed
when ligand-bound protein conformations were used as
starting points.The study demonstrated that even subtle
protein conformational changes that result from ligand
binding were sufficient to significantly influence the
quality of docking results. Nevertheless, homology
conformations within a binding site,can dramatically
influence posing oftest compounds78.Clearly,whenever
conformational changes are involved in binding,rigidly
defined binding sites are limited in their predictive poten-
tial.Finally,it has been observed that structure-based
virtual screening often selects compounds that are bio-
logically promiscuous and are therefore termed ‘frequent
hitters79,80.The fairly unspecific inhibition by such com-
pounds can,at least in part,be attributed to dominating
hydrophobic character and aggregation effects that tend
to favour their detection in both docking simulations and
screening assays (albeit for different reasons).
Selection strategies. Considering the many approxi-
mations and limitations involved in system set-up,
posing and scoring, one might ask the question of
Holo-:ligand-bound form of
form.The original definitions
referred to enzymes and
cofactors,rather than ligands,
but ligands and cofactors are
often synonymously used.
Box 5 | Search techniques
Monte Carlo algorithm in its basic form:
•Generate an initial configuration ofa ligand in an active site consisting ofa random
conformation,translation and rotation.
•Score the initial configuration.
•Generate a new configuration and score it.
•Use a Metropolis criterion (see below) to determine whether the new configuration is
•Repeat previous steps until the desired number ofconfigurations is obtained.
Ifa new solution scores better than the previous one,it is immediately accepted.Ifthe
configuration is not a new minimum,a Boltzmann-based probability function is applied.
Ifthe solution passes the probability function test,it is accepted;ifnot,the configuration
Molecular dynamics is a simulation technique that solves Newton’s equation ofmotion for
an atomic system:Fi= miai,in which Fis force,mis mass and ais acceleration.The force
on each atom is calculated from a change in potential energy (usually based on molecular
mechanics terms) between current and new positions:Fi= –(dE/ri),in which ris distance.
Atomic forces and masses are then used to determine atomic positions over series ofvery
small time steps:Fi = mi(d2ri/dt2),in which tis time.This provides a trajectory ofchanges
in atomic positions over time.Practically,it is easier to determine time-dependent atomic
positions by first calculating accelerations aifrom forces and masses,then velocities vifrom
ai= dvi/dtand,ultimately,positions from velocities vi= dri/dt.
Genetic algorithms are a class ofcomputational problem-solving approaches that adapt
the principles ofbiological competition and population dynamics.Model parameters are
encoded in a ‘chromosome’and stochastically varied.Chromosomes yield possible
solutions to a given problem and are evaluated by a fitness function.The chromosomes
that correspond to the best intermediate solutions are subjected to crossover and
mutation operations analogous to gene recombination and mutation to produce the next
generation.For docking applications,the genetic algorithm solution is an ensemble of
possible ligand conformations.
Tabu search algorithm in its basic form:
•Make nsmall random changes to the current conformation.
•Rank each change according to the value ofthe chosen fitness function.
•Determine which changes are ‘tabu’(that is,previously rejected conformations).
•Ifthe best modification has a lower value than any other accepted so far,accept it,even
ifit is in the ‘tabu’;otherwise,accept the best ‘non-tabu’change.
•Add the accepted change to the ‘tabu’list and record its score.
•Go to the first step.
NATURE REVIEWS | DRUG DISCOVERY
VOLUME 3 | NOVEMBER 2004 | 943
R E V I E W S
Hit identification.The ultimate measure of success for
the methods discussed above is their ability to produce
significant hit rates while reducing the number of com-
pounds that need to be tested.A number of successful
case studies have been published over the years that
demonstrate the potential of the approach (one also
needs to take into account that many successful applica-
tions in pharmaceutical settings are unlikely to be dis-
closed). TABLE 1 summarizes recent case studies on a
wide variety oftargets that have produced some impres-
sive results.It is interesting to note that nearly all groups
performed pre-filtering and used two-dimensional
similarity methods and shape or drug-like filters to
reduce the number of database compounds for the
time-consuming steps of flexible docking,elaborate
scoring and visual analysis.Hits with at least low-micro-
molar potency were usually found,often without biasing
search calculations towards previously identified hits.
The results also mirror the general trend that hits in the
micromolar range are much more frequently identified
than nanomolar hits in these calculations (this is similar
to the situation in biological screening).Major reasons
for this are that newly identified active compounds are
rarely optimized for potency against a given target and
that nanomolar potency is typically only obtained after
chemical modification in the course of hit-to-lead
transition and lead optimization.
Structure-based lead optimization
In addition to hit identification,docking techniques are
increasingly used to support lead optimization efforts.
Here,the scenario changes: to facilitate a hit-to-lead
transition,the compound potency typically has to be
increased by two to three orders of magnitude and rela-
tively small chemical modifications can lead to signifi-
cant changes in binding.The requirement to estimate
the effects of relatively small chemical changes further
complicates the calculations and therefore distinguishing
a micromolar compound from a nanomolar analogue
often requires much greater accuracy than typical dock-
ing and scoring can provide.However,once hits or leads
have been co-crystallized with their targets and exact
binding conformations have been established,docking
ofanalogues can be facilitated by the application ofalgo-
rithms such as ‘anchored search’24that model compound
modifications on pre-defined core fragments of leads.
These ‘conservatively’ predicted complexes usually
involve only a limited number ofanalogues,and so alter-
native and consensus scoring schemes can be easily
explored.Typical structure-based analogue design is
illustrated in FIG.5.At the very least,automated analogue
design and evaluation makes it possible to quickly elimi-
nate molecules that are too large or do not satisfy binding
constraints,and shifts focus towards more promising
synthetic candidates.For example,a series of caspase-3
inhibitors was optimized starting from a co-crystal
structure with salicylic acid88.Modelling of analogues
resulted in a compound with 20-nM potency whose
modelled structure was experimentally confirmed.
Going beyond a one-by-one evaluation of analogues,
combining docking and design of analogue libraries
models built in the presence of high sequence similarity
provided reasonable docking templates.A similar study
was undertaken to examine the value of homology
modelling within the protein kinase family76.Here,the
quality of calculated poses was found to correlate well
with ligand enrichment factors.Again,crystal structures
of ligand-bound target sites provided the best results
when used as templates. The importance of protein
conformation was recently also demonstrated by
cross-docking studies on trypsin,thrombin and HIV-1
protease69. Error rates in docking of ligands to apo-
binding sites correlated with the magnitude of protein
structural change observed as a result ofbinding.
Pre-screening: three-dimensional filtering.In addition to
conventional one/two-dimensional filters such as the
rule-of-five83,three-dimensional filter functions have
been implemented to efficiently pre-screen very large
databases and reduce the final number of docking and
scoring steps.For example,shape similarity methods can
be applied for filtering.The heuristic is based on identify-
ing similar molecular shapes on the basis of signatures,
triplets,quartets or higher-order groups of atoms84–86.
However,these shape filters are usually limited to pre-
screening ofdatabases that contain single molecular con-
formations,which can be a source of false-negatives.
In addition,pharmacophore-based screening can be
carried out where pre-defined chemical and geometric
features in compounds are matched87.Signatures and bit
strings derived from triplets of distances84and surface
triplets and histograms85have been be used to identify
preferred candidates. Recently, a ray-tracing-based
approach86has been applied to calculate shape signatures
of molecules for database searching. These types of
descriptors are also highly conformation-dependent and
therefore limited in their predictive value when only a
single (hypothetical) molecular conformation is used.
Box 6 | Types of scoring functions
944 | NOVEMBER 2004 | VOLUME 3
R E V I E W S
binding of selected functional groups.Other algorithms
that grow and score compounds within binding sites are
implemented in programs such as Groupbuild94,
GenStar95,Grow96and Growmol97.A known limitation
of such approaches is the difficulty of computationally
estimating the synthetic accessibility of‘designer’mole-
cules.However,the SYNOPSIS program represents a
recent effort to couple de novoand synthetic design98.
For example,a mini-library containing only 28 com-
pounds was selected from 373 million possible candidate
molecules in a study targeting HIV reverse transcriptase
(HIV-RT).This was achieved by use of a genetic algo-
rithm to simultaneously evaluate conformational and
synthetic parameters embedded in a fitness function.Of
the 28 selected compounds,18 could be synthesized and
ofthese molecules,10 were found to be active below 100
µM.In another study,carbonic anhydrase inhibitors
provides a particularly promising route to lead optimiza-
tion. This has been well demonstrated by a study on
cathepsin D that produced low-nM inhibitors by itera-
tive anchored docking calculations and targeted library
design89.This approach is attractive because docking and
scoring are no longer challenged to predict a ‘home run’
modification to a lead,but rather to prioritize sites and
groups for experimental modification (which is well
within the accuracy limits ofthe calculations).
De novo design.An early method for the de novodesign
of compounds in active sites is multiple-copy simulta-
neous search (MCSS)90,91.Many small fragments are
docked and simultaneously minimized within an active
site.After scoring and sorting,preferred fragments are
combined into larger molecules.Much like LUDI92,93,
results ofMCSS can provide a map oflikely sub-sites for
d van der Waals
Figure 3 | Modelling molecular recognition. a | Structure of p38 mitogen-activated protein kinase with bound inhibitor BIRB796
(PDB code: 1KV2). The inhibitor is shown with its electrostatic potential surface. b | Enlarged view of the active site. c | Close-
up view of the interaction between residue Glu71 and BIRB796. Hydrogen bonding (H-bond) and van der Waals interactions
are colour-coded red and green, respectively. d | Schematic representation of functions used to model pair-wise interactions
that contribute to binding. Interactions are calculated as a function of the distance (rij) between two atoms i and j. Left of part d:
van der Waals interaction given by a 12–6 Lennard–Jones potential (note the smoother attractive part of the potential compared to
hydrogen bond term). Middle of part d: hydrogen-bond potential given by a ‘harder’ 12–10 Lennard–Jones potential (see also BOX 2).
This term is angle-dependent (as indicated in c). Right of part d: electrostatic potential for two like (blue) or opposite (black) charges of
same magnitude calculated using a distance-dependent dielectric constant of 4r.
NATURE REVIEWS | DRUG DISCOVERY
VOLUME 3 | NOVEMBER 2004 | 945
R E V I E W S
Active-site analysis.Graphical computational analysis of
binding sites has greatly contributed to structure-based
drug design since its early days.Docking and simulation
techniques have also been applied to analyse features of
the active site, including various hydrophobic and
hydrophilic molecular fields that can identify promising
areas for ligand docking and/or de novodesign112.Surface
maps and molecular fields are mostly stored on grids that
are used to semi-quantitatively compare active sites in
homologous enzymes to explore differences in speci-
ficity112.The evaluation ofpotential interactions in active
sites can complement docking analyses.Another recent
approach generates structural interaction fingerprints
(SIFts) that allow pre-screening for potential ligands in
databases prior to docking113.In an exploration of the
active site of trypanothione reductase, 44 diverse
inhibitors were initially docked and the resulting confor-
mations were sampled and used to train a scoring func-
tion for this enzyme114.Then 2,500 novel compounds
were docked into the active site and evaluated using this
scoring scheme;13 compounds were selected for testing
and 9 were found to be active.
Active-site analysis techniques can also be applied in
combination with quantitative methods.For example,
in a study offactor Xa,QSAR models were established for
a series ofamidino inhibitors115.A total of120 analogues
were then docked into active sites using four crystal struc-
tures and the resulting molecular alignments were used
to calculate molecular fields.The resulting fields guided
the design of analogue libraries.In a similar vein,133
known factor Xa inhibitors were docked,scored and sub-
jected to regression analysis116.Scoring terms were fitted
to experimental binding energies to develop a factor Xa-
specific scoring scheme.Applying this scheme,80% of
known inhibitors could be retrieved from a compound
library with only a 15% false-positive rate116.
Absorption, distribution, metabolism and excretion
properties.Docking techniques are currently also applied
to aid in structure-based absorption,distribution,metab-
olism and excretion (ADME) evaluation.Cytochrome
were constructed using a Monte Carlo combinatorial
growth algorithm and a knowledge-based scoring
scheme75.From ~100,000 theoretical candidates,only
two compounds were selected for synthesis but both
exhibited sub-nM potency.
Simulations.Free-energy simulations are applicable to
evaluate limited numbers of analogues;for example,a
series ofthrombin inhibitors99.Various approximations
have been proposed for reducing the complexity ofper-
turbation calculations for these purposes.For example,
the OwFeg method performs a free-energy simulation
over bound and unbound states of ligands but maps
energy changes to a grid100,which greatly simplifies calcu-
lations for transforming one functional group into
another.Grid points that are energetically relevant for
various chemical modifications can be monitored during
analogue design.Moreover,linear response approxima-
tions that utilize ligand-interaction energies with the pro-
tein and solvent environment are now more commonly
applied in lead optimization101.These methods require
the availability ofat least a few experimental data points
across the range ofactivities considered to be significant.
A quantitative structure–activity relationship (QSAR) is
applied to combine non-bonded interactions that occur
within the simulated system.Linear response methods have
been shown to provide some promising results inana-
logue design studies on β-secretase102,103,HIV-RT104–107,
factor Xa108and the oestrogen receptor109.
Molecular mechanics Poisson–Boltzmann surface
area (MM/PBSA)110calculations are another molecular-
dynamics-based simulation technique involving both
force-field and solvation terms that are important for
binding.Solvation effects are estimated using a contin-
uum Poisson–Boltzmann model52.A major difference
between MM/PBSA and linear response methods is the
treatment of the ligand in its unbound state:MM/PBSA
uses NORMAL MODE analysis to calculate enthalpic and
entropic contributions to the ligand free energy.
The methodology was recently applied in analyses of
neuraminidase111and cathepsin D53inhibitors.
An oscillation in which all
particles of a system move with
the same frequency and phase.
Figure 4 | Complexity of protein–ligand interactions. The figure shows a schematic illustration of various interaction
components that need to be considered to predict the structure and binding energetics of two compounds within the same active
site. In this case, the natural cofactor of cyclic-AMP-dependent kinase (PDB code: 1atp), Mg ATP, is compared with the ATP-
binding site-directed inhibitor staurosporine. To correctly predict staurosporine as an inhibitor in a docking study, relative weights in
energy functions for the treatment of hydrophobic (indoles), hydrogen bonding (lactam ring) and ionic (aliphatic amine salt bridge,
Mg-phosphate + protein chelation) interactions must be appropriately adjusted to balance their effects. Finding preferred scoring
conditions for a specific target is a non-trivial process and often involves many trial-and-error runs.
946 | NOVEMBER 2004 | VOLUME 3
R E V I E W S
energies.For efficient compound selection,relying solely
on computed scores is currently not sufficient;experi-
ence and intuition are often still a key to success.Taking
this into account,further progress can be made in estab-
lishing more advanced scoring schemes,even if it is not
possible to develop conceptually novel scoring functions
in the near future.Importantly,scoring schemes can be
advanced by modifying molecular systems used for
benchmarking,calibrating selected functions for specific
applications,or determining the most relevant scoring
ranges. Scientific foundations for such efforts have
already been laid,as described in the following section.
The statistical analysis ofscore distributions resulting
from docking of large compound databases into differ-
ent target sites has enabled scoring ranges to be deter-
mined that are most likely to reflect ‘nonspecific’binding
events123.Similarly,docking of compound collections
into arbitrarily selected (or random) targets can provide
information about background or ‘noise’scoring levels,
regardless ofthe scoring functions that are applied.This
type of strategy has its roots in earlier investigations
designed to determine similarity measures for ligands on
the basis ofdocking against panels ofat least partly irrel-
evant receptor sites124.Compound ranking has also been
improved by the classification ofdatabases into groups of
similar molecules prior to docking and final selection of
only the best scoring representative of each group125.
Another knowledge-based approach is the use of three-
dimensional similarity information from co-crystallized
ligands as an additional constraint or scoring term126.
Scoring schemes can also be improved by tailoring
them to a specific target site, to designed sites or to
multiple related receptors.For example,altered binding
P450 isoforms are major drug-metabolizing enzymes and
have become focal points in the study of rapid metabo-
lism and drug–drug interactions117,118.Several groups
have therefore developed structure-based approaches for
the prediction ofcompounds that would be metabolized
by or inhibit P450s,and various homology models of
human P450 isoforms have been generated for these
purposes as templates for docking to predict drug
metabolism9,119–122. Recently, a crystal structure was
determined of a human P450 isoform in complex with
warfarin10. The inhibitor binds proximally to the
iron–porphyrin system in the enzyme but had no direct
interaction with the cofactor.These structural insights
should help to further refine docking studies on human
P450s and increase their predictive value.
Many of the examples and applications discussed in this
review indicate that the scoring and reliable ranking of
test compounds continue to be major bottlenecks in
structure-based virtual screening and lead optimization.
Despite a plethora ofalready available scoring functions,
further progress will be required to better account for
and balance entropic effects and electrostatic interac-
tions. Many current limitations are the result of the
assumption that implemented solvation or entropic and
electrostatic terms are generally applicable and transfer-
able to different protein systems.However,structure-
based screening calculations have produced impressive
results and many novel hits.These successes are at least
in part due to the fact that virtual-screening campaigns
mostly aim at the enrichment of active compounds,
rather than,for example,accurate calculation ofbinding
Table 1 | Comparisons of docking/scoring methods and virtual screening results
Selected docking studies focusing on structural accuracy
Comparison of 13 34 PDB structures; 100–300 compounds per activity range
Comparison of 11 100 protein–ligand complexes
41 protein–ligand complexes
1,000 kinase inhibitors
Examples of successful structure-based virtual screens
Crystal structure; 200,000 compounds; 1,000 top-scoring
clustered; rule-of-five; 15 compounds tested; 3 hits (0.4 µM)
Homology model; 250,000 NCI compounds; rigid docking
(50,000 orientations per ligand); 444 ligands flexibly docked;
13 tested; 3 hits (<100 µM)
800,000 ACD compounds reduced to 856; 9 hits; (0.25–256 µM)
Number of structuresProteins and families
genases, HIV protease
Protein kinase CK2
1,700-fold enrichment over actual HTS; 1.7 µM inhibitor through VS
ACD search, 1 hit (43 µM)
Homology model; 400,000 compounds; 12 test; one hit (80 nM)
4,000 ACD compounds; pharmacophore search; 24 tested;
12 in µM range across 6 mutants
Homology model; 200,000 NCI compounds; 35 tested; 7 hits
ACD, Available Chemicals Directory; BCL, B-cell lymphoma protein; DHFR, dihydrofolate reductase; HTS, high-throughput screening;
IGF, insulin-like growth factor; NCI, National Cancer Institute; VS, virtual screening.
NATURE REVIEWS | DRUG DISCOVERY
VOLUME 3 | NOVEMBER 2004 | 947
R E V I E W S
provides an opportunity for the de novoidentification
of active compounds,without bias towards known hits
or leads.From an algorithmic point of view,contem-
porary posing and scoring methodologies are rather
diverse.The interplay between docking and scoring
functions is fairly complex, but it is often easier to
produce reliable models of bound ligands than to dis-
tinguish ‘true’ligands from false-positives.As also
discussed in this article,further improvement ofscoring
andcompound ranking schemes does not necessarily
depend on the development of novel scoring functions.
Furthermore, compound filter functions, two- or
three-dimensional similarity-based methods and phar-
macophore models are frequently combined with
docking to reduce the number of candidate com-
pounds for fairly complex scoring calculations.
Although docking and scoring relies on many approxi-
mations,the application of these techniques during
lead optimization,often in concert with other compu-
tational methods, already extends more traditional
approaches to structure-based design.
sites that emphasize distinct chemical features can be
applied to specifically analyse or calibrate electrostatic
contributions,hydrophobic interactions or solvation
energies ofscoring functions,as has been demonstrated
in docking studies on structures ofmutated T4 lysozyme
active sites127.Moreover,relatively simple scoring terms
might be selected and refined for a specific type ofbind-
ing site128.Such tailored scoring terms can produce accu-
rate results but are,of course,not transferable.Finally,
docking against protein families is likely to improve the
predictive value of calculations focusing on single tar-
gets129and help identify specific inhibitors.This has been
illustrated,for example,by the design ofnovel antipara-
sitic agents by combining virtual screening against a target
family with structure-based compound library design130.
Docking calculations have been applied in pharmaceu-
tical research for nearly two decades.Virtual screening
on protein templates, which differs from molecular
similarity- and ligand-based virtual screening methods,
Figure 5 |Design of specific inhibitors.The active site of cyclooxygenase-2 (COX2) (PDB code: 1cx2) is shown in complex with ibu-
profen, a non-selective COX inhibitor (a), and a selective COX2 inhibitor (b); c shows a space-filling representation of the active site.
d| Several other potent COX2 inhibitors are shown. These COX2 lead compounds have different scaffolds and functional groups that can
be experimented with in the environment of the active site using docking techniques taking crystallographic information into account.
948 | NOVEMBER 2004 | VOLUME 3
R E V I E W S
1. Berman, H. M. et al.The protein data bank and the
challenge of structural genomics. Nature Struct. Biol. 7,
Westbrook, J., Feng, Z., Chen, L., Yang, H. & Berman, H. M.
The protein data bank and structural genomics. Nucleic
Acid Res. 31, 489–491 (2003).
Blundell, T. L., Jhoti, H. & Abell, C. High-throughput
crystallography for lead discovery in drug design. Nature
Rev. Drug Discov. 1, 45–54 (2002).
Bajorath, J. Integration of virtual and high-throughput
screening. Nature Rev. Drug Discov. 1, 882–894 (2002).
Walters, W. P., Stahl, M. T. & Murcko, M. A. Virtual
screening — an overview. Drug Discov. Today 3,
Langer, T. & Hoffmann, R. D. Virtual screening: an effective
tool for lead structure discovery. Curr. Pharm. Design 7,
Gohlke, H. & Klebe, G. Approaches to the description and
prediction of the binding affinity of small-molecule ligands to
macromolecular receptors. Angew. Chem. Int. Ed. 41,
A very extensive and informative review with
emphasis on quantitative analysis of protein–ligand
Kuntz, I. D., Blaney, J. M., Oatley, S. J., Langridge, R. &
Ferrin, T. E. A geometric approach to macromolecule–ligand
interactions. J. Mol. Biol. 161, 269–288 (1982).
Venhorst, J. et al.Homology modeling of rat and human
cytochrome P450 2D (CYP2D) isoforms and computational
rationalization of experimental ligand-binding specificities.
J. Med. Chem. 46, 74–86 (2003).
10. Williams, P. A. et al.Crystal structure of human cytochrome
P450 2C9 with bound warfarin. Nature 424, 464–468
11. Brooijmans, N. & Kuntz, I. D. Molecular recognition and
docking algorithms. Annu. Rev. Biophys. Biolmol. Struct.
32, 335–373 (2003).
Excellent review of research in the docking arena that
contains an instructive section on the conceptually
different processes involved in ligand–protein docking.
12. Halperin, I., Ma, B., Wolfson, H. & Nussinov, R. Principles of
docking: an overview of search algorithms and a guide to
scoring functions. Proteins 47, 409–443 (2002).
13. Burnett, R. M. & Taylor, J. S. DARWIN: a program for
docking flexible molecules. Proteins 41, 173–191 (2000).
14. Norel, R., Lin, S. L., Wolfson, H. & Nussinov, R. Shape
complementarity at protein–protein interfaces. Biopolymers
34, 933–940 (1994).
15. Norel, R., Petrey, D., Wolfson, H. & Nussinov, R.
Examination of shape complementarity in docking of
unbound proteins. Proteins 35, 403–419 (1999).
16. Connolly, M. L. Analytical molecular surface calculation.
J. Appl. Cryst. 16, 548–558 (1983).
17. Connolly, M. Solvent-accessible surface of proteins and
nucleic acids. Science 221, 709–713 (1983).
References 16 and 17 outline the theoretical
foundation of molecular surface calculations that
have also become a crucial component of many
shape-based docking algorithms.
18. Norel, R., Wolfson, H. & Nussinov, R. Small molecular
recognition: solid angles surface representation and shape
complementarity. Comb. Chem. High Throughput Screen
2, 177–191 (1999).
19. Goodford, P. J. A computational procedure for determining
energetically favorable binding sites on biologically important
macromolecules. J. Med. Chem. 28, 849–857 (1985).
This seminal paper introduced the idea of potential
energy grids and its application to understanding
protein–ligand interactions. This concept has been
applied and extended in many contemporary
20. Leach, A. R. Molecular Modelling: Principles and Applications
(Addison Wesley Longman Limited, Harlow, 1996).
21. DesJarlais, R. L. Docking flexible ligands to
macromolecular receptors by shape. J. Med Chem. 29,
22. Klebe, G. & Rarey, M. A fast flexible docking method using
an incremental construction algorithm. J. Mol. Biol. 261,
23. Kuntz, I. D. & Leach, A. R. Conformational analysis of flexible
ligands in macromolecular receptor sites. J. Comput. Chem.
13, 730–748 (1992).
24. Ewing, T. J. A., Makino, S., Skillman, A. G. & Kuntz, I. D.
DOCK 4.0: search strategies for automated molecular
docking of flexible molecule databases. J. Comput. Aided
Mol. Des. 15, 411–428 (2001).
25. Conformation search [online], <http://dock.compbio.
26. Kramer, B., Rarey, M., Lengauer, T. Evaluation of the FlexX
incremental construction algorithm for protein–ligand
docking. Proteins 37, 228–241 (1999).
27. Linnainmaa, S., Harwood, D. & Davis, L. S. Pose
determination of a three-dimensional object using triangle
pairs. IEEE Trans. Comput. Anal. Machine Intelligence 10,
An in-depth study of a computer vision technique
(pose clustering) that is utilized, for example, in FlexX.
28. Welch, W., Ruppert, J. & Jain, A. N. Hammerhead: fast, fully
automated docking of flexible ligands to protein binding
sites. Chem. Biol. 3, 449–462 (1996).
29. Kearsly, S. K., Underwood, D. J., Sheridan, R. P. &
Miller, M. D. Flexibase: a way to enhance the use of
molecular docking methods. J. Comput. Aided Mol. Des.
8, 565–582 (1994).
30. Olson, A. J. & Goodsell, D. S. Automated docking in
crystallography: analysis of the substrates of aconitase.
Proteins 17, 1–10 (1993).
31. Read, R. J. & Hart, T. N. A multiple-start Monte Carlo
docking method. Proteins 13, 206–222 (1992).
32. Dixon, J. S. & Oshiro, C. M. Flexible ligand docking using a
genetic algorithm. J. Comput. Aided Mol. Des. 9, 113–130
33. Morris, G. M., Goodsell, D. S., Halliday, R. S., Huey, R. &
Hart, W. E. Automated docking using a Lamarckian genetic
algorithm and an empirical free energy function. J. Comput.
Chem. 19, 1639–1662 (1998).
34. Jones, G., Willet, P., Glen, R. C., Leach, A. R. & Taylor, R.
Development and validation of a genetic algorithm for
flexible docking. J. Mol. Biol. 267, 727–748 (1997).
35. Westhead, D. R., Clark, D. E. & Murray, C. W. A
comparison of heuristic search algorithms for molecular
docking. J. Comput. Aided Mol. Des. 11, 209–228
36. Baxter, C. A., Murray, C. W., Clark, D. E., Westhead, D. R. &
Eldridge, M. D. Flexible docking using tabu search and an
empirical estimate of binding affinity. Proteins 33, 367–382
37. Di Nola, A., Berendsen, H. J. C. & Roccatano, D. Molecular
dynamics simulation of the docking of substrates to
proteins. Proteins 19, 174–182 (1994).
38. Trosset, J.-Y. & Scheraga, H. A. Reaching the global
minimum in docking simulations: a Monte Carlo energy
minimization approach using Bezier Splines. Proc. Natl
Acad. Sci. USA 95, 8011–8015 (1995).
39. Carlson, H. A. & McGammon, J. A. Accommodating protein
flexibility in computational drug design. Mol. Pharmacol. 57,
Informative review of approaches to treat protein
flexibility in the computational study of protein–ligand
40. Leach, A. R. Ligand docking to proteins with discrete side-
chain flexibility. J. Mol. Biol. 235, 245–356 (1994).
41. Desmet, J., Maeyer, M. D., Hazes, B. & Lasters, I. The dead
end elimination theorem and its use in protein side-chain
positioning. Nature 356, 539–542 (1992).
42. Knegtel, R. M. A., Kuntz, I. D. & Oshiro, C. M. Molecular
docking to ensembles of protein structures. J. Mol. Biol.
266, 242–440 (1997).
43. Kollman, P. A. Free energy calculations: applications to
chemical and biochemical phenomena. Chem. Rev. 93,
Review of the theory of free-energy calculations and
their areas of application, including ligand binding.
44. Simonson, T., Archontis, G. & Karplus, M. Free energy
simulations come of age: protein–ligand recognition. Acc.
Chem. Res.35, 430–437 (2002).
45. Morris, G. M. et al. Automated docking using a
Lamarckian genetic algorithm and an empirical binding
free energy function. J. Comput. Chem. 19, 1639–1662
46. Weiner, S. J., Kollman, P. A., Nguyen, D. T. & Case, D. A. An
all-atom force field for simulations of proteins and nucleic
acids. J. Comput. Chem. 7, 252 (1986).
47. Verdonk, M. L., Cole, J. C., Hartshorn, M. J., Murray, C. W.
& Taylor, R. D. Improved protein–ligand docking using
GOLD. Proteins 52, 609–623 (2003).
48. Böhm, H.-J. LUDI: rule-based automatic design of new
substituents for enzyme inhibitor leads. J. Comput. Aided
Mol. Des. 6, 593–606 (1992).
49. Eldridge, M. D., Murray, C. W., Auton, T. R., Paolini, G. V. &
Mee, R. P. Empirical scoring functions: I. The development
of a fast empirical scoring function to estimate the binding
affinity of ligands in receptor complexes. J. Comput. Aided
Mol. Des. 11, 425–445 (1997).
50. Rarey, M., Kramer, B., Lengauer, T. & Klebe, G. A fast flexible
docking method using an incremental construction
algorithm. J. Mol. Biol. 261, 470–489 (1996).
51. Rognan, D., Lauemoller, S. L., Holm, A., Buus, S. &
Tschinke, V. Predicting binding affinities of protein ligands
from three-dimensional models: application to peptide
binding to class I major histocompatibility proteins. J. Med.
Chem. 42, 4650–4658 (1999).
52. Sitkoff, D. F., Sharp, K. A. & Honig, B. Accurate calculation
of hydration free energies using macroscopic continuum
models. J. Phys. Chem. 98, 1978–1983 (1998).
53. Huo, S., Wang, J., Cieplak, P., Kollman, P. A. & Kuntz, I. D.
Molecular dynamics and free energy analyses of
cathepsin D–inhibitor interactions: insight into structure-
based ligand design. J. Med. Chem. 45, 1412–1419
54. Muegge, I. A knowledge-based scoring function for
protein–ligand interactions: probing the reference state.
Perspect. Drug Discov. Des. 20, 99–114 (2000).
55. Muegge, I. Effect of ligand volume correction on PMF
scoring. J. Comput. Chem. 22, 418–425 (2001).
56. Muegge, I. & Martin, Y. C. A general and fast scoring
function for protein-ligand interactions: a simplified potential
approach. J. Med. Chem. 42, 791–804 (1999).
57. Gohlke, H., Hendlich, M. & Klebe, G. Knowledge-based
scoring function to predict protein-ligand interactions.
J. Mol. Biol. 295, 337–356 (2000).
58. DeWitte, R. S. & Shakhnovich, E. I. SMoG: de novo design
method based on simple, fast, and accurate free energy
estimates. 1. Methodology and supporting evidence. J. Am.
Chem. Soc. 118, 11733–11744 (1996).
59. Charifson, P. S., Corkery, J. J., Murcko, M. A. & Walters, W. P.
Consensus scoring: a method for obtaining improved hit rates
from docking databases of three-dimensional structures into
proteins. J. Med. Chem. 42, 5100–5109 (1999).
This study introduced the concept of consensus
scoring as an approach to balance imperfections of
single scoring functions and improve prediction
60. Wang, R., Lai, L. & Wang, S. Further development and
validation of empirical scoring functions for structure-based
binding affinity prediction. J. Comput. Aided Mol. Des. 16,
61. Perez, C. & Ortiz, A. R. Evaluation of docking functions for
protein–ligand docking. J. Med. Chem. 44, 3768–3785 (2001).
62. Good, A. C. et al.Analysis and optimization of structure-
based virtual screening protocols 2. Examination of docked
ligand orientations sampling methodology: mapping a
pharmacophore for success. J. Mol. Graph. Model. 22,
63. Baxter, C. A. et al.New approach to molecular docking and
its application to virtual screening of chemical databases.
J. Chem. Inf. Comput. Sci. 40, 254–262 (2000).
64. GOLD Version 1.2. [online], <http://www.ccdc.cam.ac.uk/
65. Sotriffer, C. A., Gohlke, H. & Klebe, G. Docking into
knowledge-based potential fields: a comparative evaluation
of DrugScore. J. Med. Chem. 45, 1967–1970 (2002).
66. Wang, R., Lu, Y. & Wang, S. Comparative evaluation of 11
scoring functions for molecular docking. J. Med. Chem. 46,
67. McGann, M. R., Almond, H. R., Nicholls, A., Grant, J. A. &
Brown, F. K. Gaussian docking functions. Biopolymers 68,
68. Schultz-Gasch, T. & Stahl, M. Binding site characteristics in
structure-based virtual screening: evaluation of current
docking tools. J. Mol. Model 9, 47–57 (2003).
69. Erickson, J. A., Jalaie, M., Robertson, D. H., Lewis, R. A. &
Vieth, M. Lessons in molecular recognition: the effects of
ligand and protein flexibility on molecular docking accuracy.
J. Med. Chem. 47, 45–55 (2004).
70. Kontoyianni, M., McClellan, L. M. & Sokol, G. S. Evaluation
of docking performance: comparative data on docking
algorithms. J. Med. Chem. 47, 558–565 (2004).
71. Smith, R., Hubbard, R. E., Gschwend, D. A., Leach, A. R. &
Good, A. C. Analysis and optimization of structure-based
virtual screening protocols 3. New Methods and old
problems in scoring function design. J. Mol. Graph. Model.
22, 41–53 (2003).
72. Still, W. C., Tempczyk, A., Hawley, R. C. & Hendrickson, T.
Semianalytical treatment of solvation for molecular
mechanics and dynamics. J. Am. Chem. Soc. 112,
73. Ghosh, A., Rapp, C. S. & Friesner, R. A. A generalized Born
model based on a surface integral formulation. J. Phys.
Chem. B 102, 10983–10990 (1998).
74. Nissink, J. W. M. et al.A new test set for validating
predictions of protein–ligand interaction. Proteins 49,
75. Grzybowski, B. A., Ishchenko, A. V., Shimada, J. &
Shakhnovich, E. I. From knowledge-based potentials to
combinatorial lead design in silico. Acc. Chem. Res. 35,
76. Diller, D. J. & Li, Y. Kinases, homology models, and high
throughput docking. J. Med. Chem. 46, 4638–4647 (2003).
77. DesJarlais, R. L. et al.Using shape complementarity as an
initial screen in designing ligands for a receptor binding site
of known three-dimensional structure. J. Med. Chem. 31,
NATURE REVIEWS | DRUG DISCOVERY Download full-text
VOLUME 3 | NOVEMBER 2004 | 949
R E V I E W S
78. Dean, P. M. & Poornima, C. S. Hydration in drug design. 1.
Multiple hydrogen-bonding features of water molecules in
mediating protein–ligand interactions. J. Comput. Aided
Mol. Des. 9, 500–512 (1995).
79. McGovern, S. L., Caselli, E., Grigorieff, N. & Shoichet, B. K.
A common mechanism underlying promiscous inhibitors
from virtual and high-throughput screening. J. Med. Chem.
45, 1712–1722 (2002).
80. Roche, O. et al.Development of a virtual screening method
for identification of ‘frequent hitters’ in compound libraries.
J. Med. Chem. 45, 137–142 (2002).
81. Doman, T. N. et al.Molecular docking and high-throughput
screening for novel inhibitors of protein tyrosine
phosphatase-1B. J. Med. Chem. 45, 2213–2221 (2002).
An impressive example of the performance of
structure-based virtual screening.
82. McGovern, S. L. & Shoichet, B. K. Information decay in
molecular docking screens against holo, apo and modeled
conformations of enzymes. J. Med Chem. 46, 2895–2907
Informative analysis of the influence of chosen
protein-structure templates on the quality of docking
83. Lipinski, C. A. & Christopher, A. L. Experimental and
computational approaches to estimate solubility and
permeability in drug discovery and development settings.
Adv. Drug Deliv. Rev. 23, 3–25 (1997).
84. Nilakantan, R., Bauman, N. & Venkataraghavan, R. New
method for rapid characterization of molecular shapes:
applications in drug design. J. Chem. Inf. Comput. Sci. 33,
85. Good, A. C., Ewing, T. J. A., Gschwend, D. A. & Kuntz, I. D.
New molecular shape descriptors: application in database
screening. J. Comput. Aided Mol. Des. 9, 1–12 (1995).
86. Zauhar, R. J., Moyna, G., Tian, L., Li, Z. & Welsh, W. J.
Shape signatures: a new approach to computer-aided
ligand-and receptor-based drug design. J. Med. Chem. 46,
87. Rastelli, G. et al.Docking and database screening reveal
new classes of Plasmodium falciparum dihydrofolate
reductase inhibitors. J. Med. Chem. 46, 2834–2845 (2003).
88. Choong, I. C. et al.Identification of potent and selective
small-molecule inhibitors of caspase-3 through the use of
extended tethering and structure-based drug design.
J. Med. Chem. 45, 5005–5022 (2002).
89. Kick, E. K. et al.Structure-based design and combinatorial
chemistry yield low nanomolar inhibitors of cathepsin D.
Chem. Biol. 4, 297–307 (1997).
An instructive study highlighting the potential of
interfacing docking analysis and targeted library
90. Karplus, M. & Miranker, A. Functionality maps of binding
sites: a multiple copy simultaneous search method. Proteins
11, 29–34 (1991).
91. Caflisch, A. Computational combinatorial ligand design:
application to human α-thrombin. J. Comput. Aided Mol.
Des. 10, 372–396 (1996).
92. Böhm, H. J. The development of a simple empirical scoring
function to estimate the binding constant for a
protein–ligand complex of known three-dimensional
structure. J. Comput. Aided Mol. Des. 8, 243–256 (1994).
Pioneering development of an empirical scoring
function using multiple linear regression to calculate
coefficients for the most important terms
contributing to ligand binding.
93. Böhm, H. J. Prediction of binding constants of protein
ligands: a fast method for the polarization of hits obtained
from de novo design on 3D database search programs.
J. Comput. Aided Mol. Des. 12, 309–323 (1998).
94. Murcko, M. A. & Rotstein, S. H. GroupBuild: a fragment-
based method for de novo drug design. J. Med. Chem.
36, 1700–1710 (1993).
95. Murcko, M. A. & Rotstein, S. H. GenStar: a method for de
novo drug design. J. Comput. Aided Mol. Des. 7, 23–43
96. Howe, W. J. & Moon, J. B. 3D database searching and de
novo construction methods in molecular design. Comput.
Meth. 3, 697–711 (1990).
97. Bohacek, R. S. & McMartin, C. Multiple highly diverse
structures complementary to enzyme binding sites: results
of extensive application of a de novo design method
incorporating combinatorial growth. J. Am. Chem. Soc.
116, 5560–5571 (1994).
98. Vinkers, H. M. et al.SYNOPSIS: SYNthesize and OPtimize
system in silico. J. Med. Chem. 46, 2765–2773 (2003).
99. Guimaraes, C. R. W. & de Alencastro, R. B. Thrombin
inhibition by novel benzamidine derivatives: a free-energy
perturbation study. J. Med. Chem. 45, 4995–5004 (2003).
100. Pearlman, D. A. & Charifson, P. S. Improved scoring of
ligand–protein interactions using OWFEG free energy grids.
J. Med. Chem. 44, 502–511 (2001).
101. Aquist, J., Medina, C. & Samuelsson, J. E. A new method
for predicting binding affinity in computer-aided drug design.
Protein Eng. 7, 385–391 (1994).
This article presents an early formulation and use of
linear response and linear interaction approximations
in estimating binding affinity of protein ligands.
102. Tounge, B. A. & Reynolds, C. H. Calculation of the binding
affinity of β-secretase inhibitors using the linear interaction
energy method. J. Med. Chem. 46, 2074–2082 (2003).
103. Rizzo, R. C., Wang, D.-P., Tirado-Rives, J. & Jorgensen, W. L.
Validation of a model for the complex of HIV-1 reverse
transcriptase with sustiva through computation of resistance
profiles. J. Am. Chem. Soc. 122, 12898–12900 (2003).
104. Rizzo, R. C., Tirado-Rives, J. & Jorgensen, W. L. Estimation
of binding affinities for HEPT and nevirapine analogues
with HIV-1 reverse transcriptase via Monte Carlo
simulations. J. Med. Chem. 44, 145–154 (2003).
105. Kroeger-Smith, M. B. et al.Molecular modeling calculations
of HIV-1 reverse transcriptase nonnucleoside inhibitors:
correlation of binding energy with biological activity for novel
2-aryl-substituted benzimidazole analogues. J. Med. Chem.
46, 1940–1947 (2003).
106. Udier-Blagovic, M., Tirado-Rives, J. & Jorgensen, W. L.
Validation of a model for the complex of HIV-1 reverse
transcriptase with nonnucleoside inhibitor TMC125. J. Am.
Chem. Soc. 125, 6016–6017 (2003).
107. Rizzo, R. C. et al.Prediction of activity for nonnucleoside
inhibitors with HIV-1 reverse transcriptase based on Monte
Carlo simulations. J. Med. Chem. 45, 2970–2987 (2002).
108. Ostrovsky, D., Udier-Blagovic, M. & Jorgensen, W. L.
Analyses of activity for Factor Xa inhibitors based on Monte
Carlo simulations. J. Med. Chem. 46, 5691–5699 (2003).
109. van Lipzig, M. M. et al.Prediction of ligand binding affinity
and orientation of xenoestrogens to the estrogen receptor
by molecular dynamics simulations and the linear interaction
energy method. J. Med. Chem. 47, 1030 (2004).
This work provides a good example of linear
interaction methods applied to binding energies
ranging over many orders of magnitude.
110. Kollman, P. A. et al.Calculating structures and free energies
of complex molecules: combining molecular mechanics
and continuum models. Acc. Chem. Res. 33, 889–897
111. Masukawa, K. M., Kollman, P. A. & Kuntz, I. D. Investigation
of neuraminidase-substrate recognition using molecular
dynamics and free energy calculations. J. Med. Chem. 46,
112. Sheridan, R., Holloway, M. K., McGaughey, G. B., Mosley,
R. T. & Singh, S. B. A simple method for visualizing the
differences between related receptor sites. J. Mol. Graph.
Model. 21, 71–79 (2002).
113. Deng, Z., Chuaqui, C. & Singh, J. Structural interaction
fingerprint (SIFt): a novel method for analyzing three-
dimensional protein–ligand binding interactions. J. Med.
Chem. 47, 337–344 (2004).
114. Horvath, D. A virtual screening approach applied to the
search for trypanothione reductase inhibitors. J. Med.
Chem. 40, 2412–2423 (1997).
The study details many possible scoring terms for
protein–ligand complexes and is a good example of
the value of refitting parameters for a particular
protein class and series of ligands.
115. Matter, H. et al.Design and quantitative structure–activity
relationship of 3-amidinobenzyl-1H-indole-2-carboxamides
as potent, nonchiral and selective inhibitors of blood
coagulation factor Xa. J. Med. Chem. 45, 2749–2769 (2002).
116. Murcia, M. & Ortiz, A. R. Virtual screening with flexible docking
and COMBINE-based models. Application to a series of
factor Xa inhibitors. J. Med. Chem. 47, 805–820 (2004).
117. van de Waterbeemd, H. & Gifford, E. ADMET in silico
modelling: towards prediction paradise? Nature Rev. Drug
Discov. 2, 192–204 (2003).
118. Omiecinski, C. J. Concise review of the cytochrome P450s
and their roles in toxicology. Toxicol. Sci. 48, 151–156 (1999).
119. de Groot, M. J., Ackland, M. J., Horne, V. A., Alex, A. A. &
Jones, B. C. Novel approach to predicting P450-mediated
drug metabolism: development of a combined protein and
pharmacophore model for CYP2D6. J. Med. Chem. 42,
120. de Groot, M. J., Ackland, M. J., Horne, V. A., Alex, A. A. &
Jones, B. C. A novel approach to predicting P450 mediated
drug metabolism. CYP2D6 catalyzed n-dealkylation
reactions and qualitative metabolite predictions using a
combined protein and pharmacophore model for CYP2D6.
J. Med. Chem. 42, 4062–4070 (1999).
121. de Groot, M. J. Development of a combined protein and
pharmacophore model for cytochrome P450 2C9. J. Med.
Chem. 45, 1983–1993 (2002).
122. Park, J.-Y. & Harris, D. Construction and assessment of
models of CYP2E1: Predictions of metabolism from docking,
molecular dynamics, and density functional theoretical
calculations. J. Med. Chem. 46, 1645–1660 (2003).
123. Godden, J. W., Stahura, F. L. & Bajorath, J. Statistical analysis
of computational docking of large compound databases to
distinct protein binding sites. J. Comput. Chem. 20,
124. Briem, H. & Kuntz, I. D. Molecular similarity based on DOCK-
generated fingerprints. J. Med. Chem. 39, 3401–3408 (1996).
125. Su, A. I. et al.Docking molecules by families to increase the
diversity of hits in database screens: computational strategy
and experimental evaluation. Proteins42, 279–293 (2001).
126. Rognan, D., Lauemoller, S. L., Holm, A., Buus, S., Tschinke V.
Predicting binding affinities of protein ligands from three-
dimensional models: application to peptide binding to class I
major histocompatibility proteins. J. Med. Chem. 42,
127. Wei, B. Q., Baase, W. A., Weaver, L. H., Matthews, B. W. &
Shoichet, B. K. A model binding site for testing scoring
functions in molecular docking. J. Mol. Biol. 322, 339–355
128. Fradera, X., Knegtel, M. A., Mestres, J. Similarity-driven
flexible ligand docking. Proteins40, 623–626 (2000).
129. Lamb, M. L. et al.Design, docking, and evaluation of multiple
libraries against multiple targets. Proteins42, 296–318 (2001).
130. Aronov, A. M., Munagala, N. R., Kuntz, I. D. & Wang, C. C.
Virtual screening of combinatorial libraries across a gene
family in search of inhibitors of Giardia lamblia guanine
phosphoribosyltransferase. Antimicrob. Agents Chemother.
45, 2571–2576 (2001).
131. Wang, R., Liu, L., Lai, L. & Tang,Y. SCORE: a new empirical
method for estimating the binding affinity of a protein-ligand
complex. J. Mol. Model4, 379–394 (1998).
132. Tao, P. & Lai, L. Protein ligand docking based on empirical
method for binding affinity estimation. J. Comput. Aided Mol.
Des. 15, 429–446 (2001).
133. Chemical Computing Group. MOE. 2003. Montreal, Quebec,
134. Friesner, R. A. et al.Glide: a new approach for rapid, accurate
docking and scoring. 1. Method and assessment of docking
accuracy. J. Med. Chem. 47, 1739–1749 (2004).
135. Kearsley, S. K., Underwood, D. J., Sheridan, R. P. &
Miller, M. D. Flexibases: a way to enhance the use of
molecular docking methods. J. Comput. Aided Mol. Des. 8,
136. Peng, H. et al.Identification of novel inhibitors of BCR-ABL
tyrosine kinase via virtual screening. Bioorg. Med. Chem. Lett.
13, 3693–3699 (2003).
137. McNally, V. A. et al.Identification of a novel class of inhibitor
of human and Escherichia coli thymidine phosphorylase by
in silico screening. Bio. Med. Chem. Lett. 13, 3705–3709
138. Brenk, R. et al.Virtual screening for submicromolar leads of
tRNA-guanine transglycosylase based on a new unexpected
binding mode detected by crystal structure analysis. J. Med.
Chem. 46, 1133–1143 (2003).
139. Kamionka, M. et al.In silicoand NMR identification of
inhibitors of the IGF-I and IGF-Binding protein-5 interaction.
J. Med. Chem. 45, 5655–5660 (2002).
140. Vangrevelinghe, E. et al.Discovery of a potent and selective
protein kinase CK2 inhibitor by high-througput docking.
J. Med. Chem. 46, 2656–2662 (2003).
141. Enyedy, I. J. et al.Discovery of small-molecule inhibitors of
Bcl-2 through structure-based computer screening. J. Med.
Chem. 44, 4313–4324 (2001).
H.D. and J.R.F. contributed equally to this paper. This manuscript
is dedicated to Wolfram Saenger, Free University Berlin, on the
occasion of his sixty-fifth birthday.
Competing interests statement
The authors declare no competing financial interests.
Protein Structure Prediction Center:
Research Collaboratory for Structural Biology Protein Data
Biomolecular Interaction Network Database:
Drug Design Resources: http://www.drugdesign.org
Access to this interactive links box is free online.