Review

Key Topics in Molecular Docking for Drug Design

Pedro H. M. Torres 1, Ana C. R. Sodero 2, Paula Jofily 3 and Floriano P. Silva-Jr 4,*
1 Department of Biochemistry, University of Cambridge, Cambridge, CB2 1GA, UK
2 Department of Drugs and Medicines; School of Pharmacy; Federal University of Rio de Janeiro, Rio de
Janeiro 21949-900, RJ, Brazil
3 Laboratório de Modelagem e Dinâmica Molecular, Instituto de Biofísica Carlos Chagas Filho, Universidade

Federal do Rio de Janeiro, Rio de Janeiro 21949-900, RJ, Brazil

4 Laboratório de Bioquímica Experimental e Computacional de Fármacos, Instituto Oswaldo Cruz,

FIOCRUZ, Rio de Janeiro 21949-900, RJ, Brazil

* Correspondence: [email protected]; Tel.: +55-21-38658248

Received: 3 June 2019; Accepted: 10 July 2019; Published: 15 September 2019

Abstract: Molecular docking has been widely employed as a fast and inexpensive technique in the
past decades, both in academic and industrial settings. Although this discipline has now had
enough time to consolidate, many aspects remain challenging and there is still not a straightforward
and accurate route to readily pinpoint true ligands among a set of molecules, nor to identify with
precision the correct ligand conformation within the binding pocket of a given target molecule.
Nevertheless, new approaches continue to be developed and the volume of published works grows
at a rapid pace. In this review, we present an overview of the method and attempt to summarise
recent developments regarding four main aspects of molecular docking approaches: (i) the available
benchmarking sets, highlighting their advantages and caveats, (ii) the advances in consensus
methods, (iii) recent algorithms and applications using fragment-based approaches, and (iv) the use
of machine learning algorithms in molecular docking. These recent developments incrementally
contribute to an increase in accuracy and are expected, given time, and together with advances in
computing power and hardware capability, to eventually accomplish the full potential of this area.

Keywords: computer-aided drug design; structure-based drug design; benchmarking sets;

consensus methods; fragment-based; machine learning

1. Introduction
Molecular docking is a method which analyses the conformation and orientation (referred together
as the “pose”) of molecules into the binding site of a macromolecular target. Searching algorithms
generate possible poses, which are ranked by scoring functions [1]. Several software were developed
during the last decades, amongst which are some well-known examples, such as AutoDock [2]⁠,
AutoDock Vina [3], DockThor [4,5], GOLD [6,7], FlexX [8]⁠ and Molegro Virtual Docker [9].
The first step in a docking calculation is to obtain the target structure, which commonly consists
of a large biological molecule (protein, DNA or RNA) [10] (Figure 1)⁠. The structures of these
macromolecules can be readily retrieved from the Protein Data Bank (PDB) [11],⁠ which provides
access to 3D atomic coordinates obtained by experimental methods. However, it is not unusual that
the experimental 3D structure of the target is not available. In order to overcome this issue,
computational prediction methods, such as comparative and ab initio modelling can be used to obtain
the three-dimensional structure of proteins [1]⁠.

Int. J. Mol. Sci. 2019, 20, 4574; doi:10.3390/ijms20184574 www.mdpi.com/journal/ijms

Int. J. Mol. Sci. 2019, 20, 4574 2 of 29

Figure 1. General workflow of molecular docking calculations. The approaches normally start by
obtaining 3D structures of target and ligands. Then, protonation states and partial charges are
assigned. If not previously known, the target binding site is detected, or a blind docking simulation
may be performed. Molecular docking calculations are carried out in two main steps: posing and
scoring, thus generating a ranked list of possible complexes between target and ligands.

Usually, the binding site location on which to focus the docking calculations is known. However,
when the binding region information is missing, there are two commonly employed approaches:
either the most probable binding sites are algorithmically predicted or a “blind docking” simulation
is carried out. The latter has a high computational cost, since the search covers all the target structure
[12]⁠. Several available software can be used to detect binding sites. MolDock [9], for example, uses an
integrated cavity detection algorithm to identify potential binding sites. DoGSiteScorer is an
algorithm that determines possible pockets and their druggability scores, which describe the
potential of the binding site to interact with a small drug-like molecule [13]⁠. Fragment Hotspot Maps
[14] uses small molecular probes to identify surface regions in the receptor that are prone to interact
with small molecules. These predicted interaction sites can then be provided as the centre of the
sampling space.
Moreover, information derived from such hotspots or even from previous experimental
knowledge (e.g., NMR, mass spectrometry) can be used to generate distance restraints, which is
known to greatly increase protein-small molecule docking accuracy [15].
During docking calculations, a common strategy is to employ a grid representation that includes
precalculated potential energies for interaction within the target binding site [16]. This approach
speeds up the docking runs and basically consists of the discretisation of the binding site [17]⁠. Then,
at each grid point, interactions related to the Lennard–Jones and electrostatic potentials are
calculated.
Ligand structure is also required and can be obtained from small molecules databases, such as
ZINC [18] and PubChem [19]. These online databases facilitate the retrieval of a large number of
compounds for subsequent virtual screening. If not directly available, the 3D atomic coordinates of
Int. J. Mol. Sci. 2019, 20, 4574 3 of 29

these compounds can be obtained from the 2D structures (or even from simpler representation
schemes, such as SMILES) using several available software, such as ChemSketch (Advanced
Chemistry Development, Inc., Toronto, On, Canada, www.acdlabs.com, 2019), ChemDraw
(PerkinElmer Informatics), Avogadro [20] and Concord [21]. It is worth noting that for small molecule
ligands all that is needed initially is a stereochemically defined geometry with the correct relevant
protonation state, since conformations will be explored by the docking software in the context of the
target’s binding site.
Charges are usually assigned through algorithms that distribute the net charge of a molecule
among its constituent atoms as partial atom-centred charges. Furthermore, most docking methods
assume that a particular protonation state and charge distribution in the molecules do not change
between their bound and unbound states [3].⁠ Nevertheless, it is crucial for successful docking to
evaluate free torsions, protonation states and charge assignments. The protonation states of the
target’s amino acid residues can be critical to ligand interactions and, consequently, to the binding
affinity prediction. There are several software available to evaluate the pKa of the amino acid
residues, such as PropKa [22] and H++ [23].
Ligand protonation is also important since it affects the net charge of the molecule and the partial
charges of individual atoms. Nonetheless, each docking program will employ a different charge
assignment protocol [1]. For example, in the MolDock program, the protein and the ligands are
automatically prepared (charges and protonation states assigned) and simplified charge and
protonation schemes are used, as described by Thomsen and Christensen (2006). AutoDock uses
Gasteiger–Marsili atomic charges whereas the closely-related AutoDock Vina does not require the
assignment of atomic charges, since the terms that compose its scoring function are charge-
independent [3,24]⁠. The DockThor algorithm, as implemented in the homonymous web portal,
automatically generates the topology files (i.e., atom types and partial charges) for the protein, ligand
and cofactors according to the MMFF94S force field [4,5,25]⁠.
Two aspects are crucial to docking programs: search algorithms and scoring functions. The
search algorithm analyses and generates ligand pose at a target’s binding site, taking into
consideration the roto-translational and internal degrees of freedom of the ligand [10]⁠.
Search strategies are often classified as systematic, stochastic or deterministic [16]⁠. Systematic
search algorithms explore each ligand’s degree of freedom incrementally. As the number of free
rotatable bonds increases, the number of evaluations can undergo a combinatorial explosion
[16,26,27]. This class of search algorithms can be subdivided in exhaustive, incremental construction
(which relies on the fragmentation of the ligand) and conformational ensemble [26]⁠. FlexX [8] and
eHits [28], for example, employ fragment-based approaches with systematic algorithms (incremental
construction and graph matching, respectively).
A number of algorithms were also developed to use information from protein and ligand
pharmacophores. Those algorithms try to match the distances between each of the ligand’s and
protein’s pharmacophoric points [29]. The software FLEXX-PHARM, for example, is an extended
version of FLEXX and applies pharmacophoric features as constraints into a docking calculation [30].
Stochastic search algorithms perform random changes in the ligand’s degrees of freedom.
However, this kind of algorithm does not guarantee convergence to the best solution. To improve it,
an iterative process can be performed. Monte Carlo, Evolutionary Algorithms (including genetic),
Tabu Search and Swarm Optimisation are some of the most common stochastic algorithm
implementations [26]⁠. Several software use stochastic algorithms as search methods, such as
AutoDock [2], GOLD [6]⁠, DockThor [4,5,25]⁠ and MolDock [9]⁠ (Table 1).
Int. J. Mol. Sci. 2019, 20, 4574 4 of 29

Table 1. Molecular docking software.

Software Posing Scoring Availability Reference

Iterated Local Search + Empirical/Knowledge- Free (Apache
Vina Trott, 2010 [3]
BFGS Local Optimiser Based License)
Lamarckian Genetic
Morris, 2009;
Algorithm, Genetic Free (GNU
AutoDock4 Semiempirical Huey, 2007
Algorithm or License)
[31,32]
Simulated Annealing
Differential Evolution
(Alternatively Simplex Thomsen,
Molegro/MolDock Semiempirical Commercial
Evolution and Iterated 2006 [9]
Simplex)
Monte Carlo stochastic
Free (GNU Koes, 2013
Smina sampling + local Empirical (customisable)
License) [33]
optimisation
Korb, 2007;
Ant Colony Academic
Plants Empirical Korb, 2009
Optimisation License
[34,35]
Abagyan,
Biased Probability
1993;
ICM Monte Carlo + Local Physics-Based Commercial
Abagyan, 1994
Optimisation
[36,37]
Systematic search +
Optimisation (XP mode Friesner, 2004
Glide Empirical Commercial
also uses anchor-and- [38]
grow)
Fragmentation and
alignment to idealised Jain, 2003; Jain
Surflex Empirical Commercial
molecule (Protomol) + 2007 [39,40]
BFGS optimisation
Physics-based (GoldScore),
Jones, 1997;
Empirical (ChemScore,
GOLD Genetic Algorithm Commercial Verdonk 2003
ChemPLP) and
[6,7]
Knowledge-based (ASP)
Free (for non-
Generic Evolutionary Empirical (includes Yang, 2004
GEMDOCK commercial
Algorithm pharmacophore potential) [41]
research)
Anchor-and-grow
Physics-based (several Academic Allen, 2015
Dock6 incremental
other options) License [42]
construction
Entropy-based multi-
GAsDock population genetic Physics-based * Li, 2004 [43]
algorithm
Fragment-Based
Rarey, 1996;
Pattern-recognition
FlexX Empirical Commercial Rarey, 1996b
(Pose Clustering) +
[8,44]
Incremental Growth
Conformer generation
Empirical (defaults to McGann, 2011
Fred + Systematic rigid body Commercial
Chemgauss3) [45]
search
Steady-state genetic
algorithm (with
Free De Magalhães,
DockThor Dynamic Modified Physics-based + Empirical
(Webserver) 2014[4,25]
Restricted Tournament
Selection method)
*Availability is unclear.
Int. J. Mol. Sci. 2019, 20, 4574 5 of 29

In deterministic search, the orientation and conformation of the ligand in each iteration is
determined by the previous state, and the new state has equal or lower energy value than the
previous one [16,26]⁠. However, this kind of algorithm has higher computational cost and often leads
to the undesired trapping of the resulting conformations to a local energy minimum [16]⁠. Examples
are energy minimisation methods and molecular dynamics (MD) simulations.
The overall size of the ligand, especially if it contains a large number of rotatable bonds impacts
most docking algorithms in a negative way, both in terms of computational cost of each individual
docking run and in terms of docking accuracy [46]. That is the case because each new rotatable bond
inherently increases the ligand’s degrees of freedom, thus increasing the number of possible
conformations. The enhanced conformational space is therefore much more complex to explore,
rendering less accurate results, usually even with increased sampling steps. The magnitude of this
effect is distinct for different algorithms [3,47] and fragment-based ones seem to exhibit superior
performance in such cases [46].
Some algorithms can combine different search strategies, and often MD simulations are used to
analyse the time-resolved trajectory of the ligand-bound system and to further pinpoint the best
docking solutions [48–51]⁠.
After the generation of thousands of ligand orientations, additional scoring functions may be
used to rank the conformations. They may be based on binding energy, free energy, or a qualitative
numerical measure to approximate interaction energies [52]⁠. Currently, scoring functions are
grouped into three major types: force field, empirical and knowledge-based [26,27,53]⁠.
Force field-based functions consist of a sum of energy terms [26]. The potential energy usually
accounts for bonded (bond length, angle, dihedrals) and nonbonded (van der Waals, electrostatic)
terms. This type of function usually neglects solvent effects and entropies [16]⁠. The DockThor
program [4], for example, employs a scoring function for pose prediction based on the MMFF94S
force field composed of three energy terms [54], i.e., the torsional term for bonded interactions, the
electrostatic potential and the Buffered-14-7 term for the van der Waals potential (Equation (1)):

𝐸 = 0.5 𝑉 (1 + 𝑐𝑜𝑠 𝛷) + 𝑉 (1 − 𝑐𝑜𝑠 2𝛷) + 𝑉 (1 + 𝑐𝑜𝑠 3𝛷)

332.0716𝑞 𝑞 (1 + 𝛿 )𝑅 ∗ (1 + 𝛾)𝑅 ∗ (1)
+ + 𝜀 −2 ,
𝜀 𝑅 +𝛿 𝑅 +𝛿 𝑅∗ 𝑅 + 𝛾𝑅 ∗

where V1, V2 and V3 are constants dependent on the types of the atoms i and j, ϕ is the i-j-k-l torsion
angle, qi and qj are the partial charges of atoms i and j, ε is the dielectric constant given by a distance-
dependent sigmoidal dielectric function [55], Rij is the internuclear separation between atoms i and j,
and δelec is the electrostatic buffering. Repulsion at short distances and van der Waals interactions are
calculated by the last term, the Buf-14-7 potential [56]. In this term, 𝜀 is the well depth, 𝑅∗ is the
minimum-energy separation (Å) that depends on the MMFF94S types of the atoms i and j, and
𝛿 = 0.07 and 𝛾 = 0.12 are the buffering constants.
Empirical scoring functions are derived from quantitative structure-activity relationships which
were first idealised by Hansh and Fujita [16,57]⁠. The goal is to predict binding affinity with high
accuracy by using known experimental binding affinity data [26]⁠. ChemScore [58] and GlideScore
[59] are examples of empirical scoring functions.
Knowledge-based functions are based on frequency of atom pairs interactions observed in
experimentally determined 3D structures of ligand-target complexes [16,26]⁠. DrugScore of FlexX
program [60] and PMF [61] are examples of knowledge-based functions.
Binding affinity prediction is still a major challenge for docking programs and most approaches
rely upon consensus scoring schemes and rescoring approaches [16,26,27]. Consensus scoring for
improving molecular docking accuracy is an ever-evolving research topic and will be addressed
further in this review.

1.1. Molecular Docking in Drug Design

Int. J. Mol. Sci. 2019, 20, 4574 6 of 29

Molecular docking is a key component of the Computer-aided Drug Design toolbox. It is part of
the so-called “structure-based drug design” methods and was first developed in the middle 80s
through early 90s for predicting the binding mode of known active compounds and virtually
screening large digital compound libraries to reduce costs and speed up drug discovery [62]. Docking
tools have also been used in the hit-to-lead optimisation process. The latter application imposes the
biggest challenge as predicting relative binding affinities for a series of related compounds has been
the Achilles heel of most docking software since the very beginning of their development.
Nevertheless, docking can still be used in hit-to-lead optimisation by indicating if the designed
analogues of a hit compound present improved molecular interactions with the target.
Another widely known shortcoming of traditional docking methodologies is the poor modelling
of receptor flexibility [63–65]. Some docking algorithms are able to partially mitigate this issue by
allowing side-chain movement of active-site residues. Nevertheless, larger conformational changes
might be triggered upon ligand-binding or might be a prerequisite to the binding event itself. A
strategy, usually referred to as Receptor Ensemble Docking (or simply Ensemble Docking) is the most
frequently used to model those scenarios. It is based on the concept of Conformational Selection and
consists in using multiple conformations of the receptor molecule, that can be obtained via different
methods, such as MD simulations [66,67], Normal Mode Analysis [68], and even by using alternative
experimentally-determined receptor conformations [69]. It is worth noting that some software, such
as GOLD and Glide have implemented functionality to execute this type of analysis.
The main limitations and challenges in the docking methodology have been identified nearly
two decades ago [16] but they are still the subject of a very active research field. As described earlier
here, two key components of the docking methodology are the conformational search algorithm and
the scoring function. The former can suffer dramatically in performance when dealing with longer
and flexible ligands, especially for shallow and chemically featureless binding sites, such as in
polymer binding proteins (e.g., peptidases and glycosidases). Force-field based scoring functions
suffer from the inherent problem of calculating binding affinities from the simplified interaction
energies necessary to keep the docking calculations fast enough to process large compound libraries.
Although binding affinities can be more accurately predicted from calculated binding free energies
the latter also suffers from a problem of subtraction of large numbers (interaction energy between the
ligand and protein on one hand and the cost of bringing the two molecules out of solvent and into an
intimate complex on the other hand), which are often calculated with sub-optimal accuracy, and yield
a small number as a result of the calculation [70].
In the following sections, we will review and discuss a selection of the main topics in the
literature for molecular docking in drug design, all of which intend to address the above discussed
limitations and advances in the methodology.

2. Benchmarking Sets
When using computational methods for molecular docking, it is paramount to assess the
performance and accuracy of the programs to be employed. This not only allows one to know the
degree of credibility that can be expected in the results, but also helps choosing the method or
program better suited to the task at hand. To that end, there are many benchmarking databases that
provide targets and ligands for docking, along with additional information such as true binding
affinity, experimental binding pose, and actives/inactives distinction. Experimental information can
then be compared to the docking program’s predictions through different statistical metrics, which
allows the assessment of its performance.

2.1. Benchmarking Sets for Pose Prediction and Binding Affinity Calculations
The development of either empirical parametric or nonparametric regression models for
docking pose and binding affinity predictions must be based on experimental data so that their
functions may be properly parameterised (or inferred) and thus better represent reality. Moreover,
the performance of these models must also be evaluated on such data. In light of this demand, there
are many benchmarking datasets which aim to group as much high-quality data as possible [71–74].
Int. J. Mol. Sci. 2019, 20, 4574 7 of 29

The most widely employed of these is PDBBind [71]. This database is a result of an effort to
screen the entire Protein Data Bank (PDB) [11] for experimentally determined 3D structures of
protein-ligand complexes and collect their experimentally measured binding affinities. There is also
a refined set of complexes [75] and a core set derived from it [76], which has become the standard set
for benchmarking scoring functions (SFs). It is noteworthy that PDBBind is also widely used in
training machine learning SFs for binding affinity predictions [77–79].
There are also benchmarking databases which encompass specific complexes or purposes, such
as protein-protein complexes [80], membrane protein-protein complexes [81], and a blind set based
on PDBBind for testing machine learning SFs [82].
Accuracy for pose prediction can be assessed by root mean square deviation (RMSD)
calculations comparing predicted pose and experimental pose. To compare binding affinity
predictions with experimentally determined affinities for a set of multiple data points, one can too
calculate RMSD for the values, but also the Pearson correlation coefficient (Rp) and the Spearman
rank-correlation (Rs) [83].

2.2. Benchmarking Sets for Virtual Screening

Benchmarking databases for virtual screening (VS) consist on datasets with selected known
active ligands and inactive decoys for a single protein target [84]. Since information on inactive
molecules is scarce in comparison to active ones, most decoys are not selected based on experimental
data but are instead putative inactive compounds [85], whose selection must be made carefully so as
to avoid artificial enrichment [86]. This scarcity occurs because active molecules are better described
and documented, however, the opposite asymmetry is observed in nature: from a varied set of
molecules which come in contact with a given protein, only a few specific ones will be active against
it. Therefore, VS programs must be capable of identifying active compounds amidst a large pool of
inactive ones, thus, benchmarking sets mirror this natural asymmetry by providing many putative
decoys for a single known active molecule. In order to prevent bias, the active and decoys sets’
characteristics must be equally balanced: one set must not be more structurally complex or diverse
than the other [87,88]; both sets should not cover small chemical spaces [84]; and there must not be
any actual binders among the decoys (Latent Actives in the Decoy Set, LADS) [89]. Datasets are
therefore curated in order to avoid bias as well as provide as much useful data as possible; the most
widely used are described as follows.
The Directory of Useful Decoys (DUD) was created based on the principle that decoys must
resemble the physical properties of the actives but be sufficiently chemically distinct to be in fact
nonbinders [90]. DUD then became the gold standard benchmark for VS [91]. It was later improved
into the Directory of Useful Decoys-Enhanced (DUD-E) [92], which selects decoys based on more
physicochemical properties, adds more targets, and provides a tool for decoy generation based on
user-input actives.
The Demanding Evaluation Kits for Objective in Silico Screening (DEKOIS) [89] was created
with special attention to avoiding poorly embedded actives and LADS. A new version, DEKOIS 2.0
[93], was released two years later with additional physicochemical properties for matching decoys
and an enhanced elimination of LADS.
The Maximum Unbiased Validation (MUV) [94] datasets were curated with special care for the
chemical diversity of the actives set, in order to avoid over-representation of chemical entities and
thus avert overestimation of performance. An exclusion of potentially unspecific active compounds
was also implemented, as well as removal of actives devoid of decoys in its chemical space.
There are also databases for assessing virtual screening with specific targets: G-Protein-Coupled
Receptor (GPCR) Ligand Library (GLL) and GPCR Decoy Database (GDD) [95], NRLiSt BDB for
nuclear receptors [96] and MUBD-HDACs for histone deacetylases [97].
It is noteworthy that it is also possible to generate decoys for specific compounds when the target
of interest is not available. User-input ligands must be provided in SMILES format, and a decoy set
is curated based on their molecular properties. DecoyFinder [98] was the first application to provide
this tool, searching the ZINC database for molecules similar to actives by comparing chemical
Int. J. Mol. Sci. 2019, 20, 4574 8 of 29

descriptors. At about the same time DecoyFinder was published, DUD was upgraded to DUD-E,
which also allows searching the ZINC database for decoys utilising the same search method
employed to construct the database’s new target subsets. In 2017, Wang et al. [99] argued that these
tools lacked computational speed for large active sets and flexible input options to avoid bias in the
user-specified active set. To address these issues, they created RADER (RApid DEcoy Retriever),
which selects decoys from four different databases, including ZINC.

2.3. Evaluation Metrics

The most widely used metrics to assess ranking performance in VS are receiver operating
characteristic (ROC) curves and enrichment factors (EF). The ROC method plots the rank’s specificity
and sensitivity into a curve whose area (area under the curve, AUC) ranges from 0 (worst
performance) to 1 (best performance), where 0.5 reflects a randomly distributed ranking order. The
calculations are made based on cut-offs throughout the whole rank, and therefore ROC reflects only
overall performance [100,101]. However, when evaluating VS performance, the enrichment at the top
of the rank is most important (i.e., the early recognition problem), since there can be found the
molecules identified by the SF as the most probable actives [102]. EF can be used to calculate the
enrichment at an early single cut-off [83] or at many cut-offs [101], which addresses the early
recognition problem, however its main setback resides in the fact that its maximum value depends
on the active/inactive ratio on the dataset [101,103].
It is noteworthy that by calculating Youden’s index (sensitivity + specificity − 1) for all cut-offs
made in the ROC curve, one can determine the optimal threshold (i.e., the cut-off with the highest
index) through which continuous binding predictions of a particular SF can be converted into to
binary active/inactive classification [104].
Other metrics have been suggested and applied to better address the early recognition problem.
For instance, the Robust Initial Enhancement (RIE) metric [105] applies weight to the active
molecules. The active will weigh closer to 1 the better ranked it is, and its weight will fall as the rank
increases. A RIE value of 1 indicates a random distribution of the rank, and its maximum value
depends on the active/inactive ratio, similarly to EF. The Boltzmann-Enhanced Discrimination of
Receiver Operating Characteristic (BEDROC) [102] incorporates the RIE weighing strategy into ROC
curves: performance is measured in a 0 to 1 range and advantage is given to better ranked actives.
One drawback of the BEDROC approach is that the magnitude of this advantage is controlled by a
single parameter, which can frustrate performance comparisons between different studies [103].
No single benchmarking set or metric can be considered to be best overall for molecular docking.
Rather, they are chosen differently depending on the inquiry, as well as carefully, in order to avoid
biasing issues. Erroneous estimations of performance negatively impact studies and are also very
hard to detect based on benchmarking results alone. Nonetheless, benchmarking datasets provide
invaluable means for quality assessment of computational methods in drug discovery.

3. Consensus Methods
With the continued development of new scoring functions (SFs) and the improvement of well-
established ones, the use of docking strategies that combine two or more SFs has become increasingly
common. That is especially interesting because the various available functions perform differently
across the spectrum of potential interactions, and presumably, in an ideal combination, the
shortcomings of a particular function may be compensated by the others.
This strategy was first suggested by Charifson and co-workers in a study in which they
benchmarked several SFs, both individually and in combination, using p38, IMPDH and HIV
protease as model systems. Their approach involved taking the intersection of the top-scoring
molecules according to two or three different functions available at the time and they found it
provided a “dramatic reduction in the number of false positives identified by individual SFs” [106].
A consensus-docking protocol will generally differ in three major aspects: (i) the means by which
the poses are obtained, (ii) the selection of the SFs, and (iii) the algorithm used to achieve the
consensus. Realistically, the number of possible procedures is overwhelming, and, to date, no single
Int. J. Mol. Sci. 2019, 20, 4574 9 of 29

protocol has been proven remarkably superior to the others. Nevertheless, it is absolutely clear that
consensus methods perform consistently better when compared to individual SFs (c.f. referenced
papers in Table 2).
The theoretical rationale for this was explored in 2001, soon after the first approaches, in a work
in which the authors simulated an idealised computer experiment where scores were generated for
a hypothetical set of 5000 compounds and the effects of consensus strategies were evaluated. The
authors suggest that the improvement is largely due to the fact that the mean value of repeated
samplings tends to be closer to the true value than any single sampling [107].
Although some initiatives have been explored to come up with composite scoring schemes that
are applied simultaneously during the posing procedure [108], in most cases, the consensus is
achieved after the conformational sampling. Moreover, it is widely accepted that the conformational
sampling is not the major bottleneck in the docking process [109,110] therefore, a greater fraction of
the developed methods generate the docking poses using a single algorithm and subsequently use a
different set of SFs to re-assess them (Tables 2 and 3). Nevertheless, several groups have focused on
obtaining more reliable poses, for example, Ren and co-workers have explored the effects of using
multiple software in the pose generation step [111]. They used a RMSD-based criterion to come up
with a representative pose derived from a minimum of three and a maximum of 11 docking
programs. A pose representative was selected for all possible combinations and their method
achieved an increase in the success rate (pose-to-reference RMSD < 2.0 Å) of approximately 5% when
compared to the best independent program.
Additionally, the concept of “consensus level” has been explored in recent works [112,113], and
similarly to the previously described approach, it uses a combination of docking software to generate
ligand poses, which are then clustered and the number of software that predict the same pose is taken
as the consensus level. This metric can then be used to reject compounds that fail to attain a certain
level and true ligands are less likely to being rejected, which, in turn increases the enrichment factors.
Another consensus posing strategy is to reject a given pose if it two or more programs fail to
“converge” to that conformation. Houston and Walkinshaw have demonstrated that the success rate
can be increased from ~60% to ~80% by simply rejecting a molecule if the RMSD between the poses
calculated by two programs (AutoDock and VINA) is greater than 2.0 Å. The idea behind this
approach is that a correct pose is more likely to be predicted by more than one algorithm, thus
eliminating the misleading orientations (which could be considered false positives) [114].
Some initiatives combine consensus posing and scoring, as is the case of the VoteDock approach
(and two correlated functions), proposed by Plewczynski et al., in which they combine cross-software
pose conformation agreement, in the form of a voting system, with a composite scoring obtained via
multivariate linear regression with results performing consistently better than individual SFs [115].
Besides consensus posing, many groups have focused their efforts on creating consensus scoring
schemes. Very recently, Perez-Castillo and co-workers have applied the Genetic Algorithm to devise
the best combination from a total of 15 SFs (or 87 scoring components) that maximises either the
enrichment factor or the BEDROC value. Their results suggest that combining scoring components,
instead of SFs themselves is a more effective strategy. Their algorithm, CompScore, is made available
as a webserver [116].
Other reported strategies for achieving scoring function consensus are sequential docking
[117,118], linear regression [119], rank-by-rank, rank-by-number, rank-by-vote [86,107,120] and
standard deviation consensus [121]. Combinations of consensus docking strategies and ligand-based
approaches have also been suggested [122,123].
Int. J. Mol. Sci. 2019, 20, 4574 10 of 29

Table 2. Consensus docking methods.

Source Ta Posingb Fc Consensus Strategy Analysis Ref.

Rank/Score
Standard Deviation Consensus
DUD-E/ curves Chaput, 2016
102/3 4 4 (SDC),
PDB Hit recovery [121]
Variable SDC (vSDC)
count
EF, ROCAUC Ericksen,
DUD-E 21 8 8 Gradient Boosting
2017 [124]
PDBBind Vina, Compound rejection if pose Houston,
228/1 2 Success rate
DUD AutoDock RMSD > 2.0 Å 2013 [114]
Multi-Objective Scoring Kang, 2019
PDB 3 GAsDock 2 EF
Function Optimisation [108]
mTORd
1 Glide 26 Linear Combination BEI Correlation Li, 2018 [119]
Inhibitors
Compression Oda, 2006
PDB 220 FlexX 9 Severale
and Accuracy [120]
Perez-
Genetic Algorithm used to
DUD-E 102 Dock 3.6 15 EF, BEDROC Castillo,
combine SF components
2019 [116]
RMSD-based pose consensus, Plewczynski,
PDBBind 1300 7 7 Success rate
multivariate linear regression 2011 [115]
Compound rejection based on Poli, 2016
DUD 35 10 10 EF
RMSD consensus level [112]
Selection of representative pose Ren, 2018
PDBBind 3535 11 11 Success rate
with minimum RMSD [111]
Supervised Learning (Random
Average RMSD, Teramoto,
PDB 100 AutoDock 11 Forests),
Success rate 2007 [125]
Rank-by-rank
PDB Compound rejection based on Tuccinardi
130/3 10 10 EF, ROCAUC
DUD RMSD consensus level (2014) [113]
PDBBind Top pose /Top Wang, 2013
421 Glide 7 Support Vector Rank Regression
CSAR Rank [126]
Rank/Score
GEMDOCK Rank-by-rank, Yang, 2005
PDB 4 2 curve, GH Score,
GOLD Rank-by-score [127]
CS index
aTotal number of targets used in the assay; bPosing software used. If more than two software were
used, than only the number is indicated; cNumber of scoring functions used; dIn this study, the dataset
was composed of 25 mammalian target of rapamycin (mTOR) kinase inhibitors retrieved from the
literature and six mTOR crystal structures retrieved from PDB; eThe purpose of this study was to
evaluate several different consensus strategies (e.g., rank-by-vote, rank-by-number, etc).
Int. J. Mol. Sci. 2019, 20, 4574 11 of 29

Table 3. Recent works using consensus docking approaches.

Best
Consensus
Target Lig. Posing Fa Hits/Test Activity Ref.
Strategy
(IC50)
3.57 × VINA, Sequential Onawole,
EBOV Glycoprotein 2 - -
107 FlexX Docking 2018 [117]
Z-scaled rank-by-
number
Liu, 2012
β-secretase (BACE1) 1.13e5 Surflex 12 Principal 2/20 51.6 μM
[128]
Component
Analysis
Sequential
Docking
Aliebrahimi,
c-Met Kinase 738 2 2 Compound - -
2017 [118]
rejection if pose
RMSD > 2.0 Å
Mokrani,
Acetylcholinesterase 14,758 4 4 vSDC[121] 12/14 47.3 nM
2019 [129]
Compound
rejection based on 13.4 μM Spena, 2019
PIN1 32,500 10 10 1/10
RMSD consensus 53.9 μMc [130]
level
Support Vector Zhan, 2014
Akt1 47 LigandFit 5 6/6b 7.7 nM
Regression [123]
Compound
Monoacylglycerol 4.80 × rejection based on Mouawad,
4 4 1/3 6.1 μM
Lipase (MAGL) 105 RMSD consensus 2019 [131]
level
aNumber of scoring functions used; bThis work consisted of a Quantitative Structure-Activity
Relationship (QSAR) model using consensus docking as descriptors. Six compounds were designed,
synthesised and tested, exhibiting IC50 values between 7.7 nM and 4.3 μM; cFirst IC50 value: inhibitory
activity against PIN1 isomerisation. Second IC50 value: inhibitory effects on ovarian cancer cell lines.

Machine learning algorithms have also been employed in the determination of the consensus in
recent developments. Early efforts used Random Forest algorithms to achieve consensus for 11
different SFs, outperforming the regular rank-by-rank approach in about 5%–10% and individual SFs
by a far greater margin [125]. Support Vector Rank Regression (SVRR) has been suggested as a
possible tool to combine seven distinct SFs (Glide- Score, EmodelScore, EnergyScore, GoldScore,
ChemScore, ASPScore and PLPScore) computed using GLIDE and GOLD docking programs, and
was shown to improve correct top pose prediction (RMSD < 2.0 Å) by 12.1% and correct top ligand
selection by 46.3% [126]. In another study, Ericksen and collaborators used gradient boosting to
derive a consensus score and benchmarked this approach using 21 targets selected from DUD-E,
gradient boosting was shown to outperform traditional consensus methods (maximum, median and
mean scores) and as well as the mean-variance consensus [124]. A summary of the aforementioned
works can be found in Table 2.
Although molecular docking was first applied over three decades ago, it is apparent, given the
virtually endless protocols, that there is still much improvement to be made in the field. In this sense,
initiatives such as the Community Structure-Activity Resource (CSAR active from 2010 to 2014)
[73,132] and the Drug Design Data Resource (D3R) [133,134] are invaluable as they promote the
standardisation of validation datasets and metrics, as well as serve as a repository for the knowledge
accumulated in the field.
A simple comparison made with a keyword search software in the SCOPUS database for the
years 1995 until 2018 (“TITLE-ABS-KEY (software AND docking) AND PUBYEAR > 1994 AND
PUBYEAR < 2019” where the word software is replaced by several of the mostly employed docking
programs) shows the relative prevalence of these software. Substituting the term software for
Int. J. Mol. Sci. 2019, 20, 4574 12 of 29

consensus, shows that consensus methods, in spite of consistently showing superior results, are less
frequently mentioned in the literature than some of the more common docking programs (at least in
the searched fields, i.e., title, abstract and keywords) (Figure 2). While one could argue that this could
be due to the fact that the fraction of works that indeed use consensus methods also mention other
software, Figure 3, which contains the ratio of (research and conference) papers mentioning
“molecular docking” OR “ligand docking” to the ones mentioning (“molecular docking” OR “ligand
docking”) AND consensus, shows that the discrepancy is even more pronounced (an average of 88.36
works that cite molecular docking per each work that mentions the word consensus—Figure 3).

Figure 2. Scopus search results for the query “TITLE-ABS-KEY (software AND docking) AND
PUBYEAR > 1994 AND PUBYEAR < 2019” where the word software is substituted for one of the eight
most common docking software or by the word consensus.

Figure 3. Ratio of the numbers of papers containing either the expression “molecular docking” or
“ligand docking” to the number of papers containing either of the two expressions AND the word
consensus.

There is also clear disparity in the level of elaborateness between the protocols used by the
groups that develop and the ones that implement these methods. As a result, the virtual screening
protocols used by the latter (such as sequential docking, rank-by-number and RMSD-based pose
rejection) are often less involved than the ones suggested by the former. Table 3 summarises recent
works that employed consensus docking in their screening methodologies, along with the best
experimentally-determined activity. Despite using more straightforward methodologies to achieve
consensus, these studies show the importance of combining distinct SFs, since they have still been
able to find relatively potent ligands. It appears that, easy-to-use, carefully designed and validated
Int. J. Mol. Sci. 2019, 20, 4574 13 of 29

docking pipelines which include consensus posing and/or scoring are called for and could be widely
adopted in structure-based drug design studies, both in academic and industrial settings.

4. Efficient Exploration of Chemical Space: Fragment-Based Approaches

4.1. The Chemical Space

Since it was first described in the late 1990s [135], fragment-based drug (or, less frequently, lead
[136]) discovery (FBD/LD) has gained a lot of attention and many drug candidates developed with
the use of such approaches have reached clinical trials [137]. The fundamental aspect that fosters its
popularity is that it allows an efficient exploration of the chemical space with relatively small
sampling, i.e., by combining smaller fragments that show high ligand efficiency, it is possible to
design very potent ligands which would, otherwise, be dispersed in a vast pool of possible molecules.
Additionally, it has been demonstrated that the probability of a given interaction between a given
ligand and receptor is inversely proportional to the ligand complexity [138,139], suggesting that
higher hit-rates could be achieved by screening less complex molecules.
In 2007, researchers from Reymond’s group at the University of Berne used a graph-based
approach to generate all possible topologies for chemically-stable compounds presenting up to 11
atoms, and they generated a database containing near 26.4 million (2.64 × 107) molecules (GDB11)
[140]. Since then, they have created new sets of increasingly larger molecules, namely containing up
to 13 heavy atoms (GDB13—9.7 × 108 molecules) and up to 17 heavy atoms (GDB17—1.66 × 1011).
These numbers might seem overwhelming, but not if compared to an astonishing 1060 estimated
drug-like molecules (with up to 30 heavy atoms) [62].
Very recently, researchers from UCSF (University of California, San Francisco) have completed
ultra-large campaigns, screening approximately 99 million compounds for AmpC β-lactamase and
138 million compounds for the D4 dopamine receptor, ultimately finding 30 compounds with sub-
micromolar activity, including one with picomolar activity (180 pM) [141]. Endeavours of such
magnitude have not been customarily undertaken, since they require great use of computational
resources, therefore, fragment-based approaches can help efficiently explore the chemical space since
(i) they have a small amount of degrees of freedom, leading to faster spatial sampling, (ii) they can
be combined to create larger, more potent ligands, requiring reduced screening libraries to achieve
comparable chemical space coverage, and (iii) the reduced complexity of fragments should lead to
increased hit-rates.
Experimentally, due to the reduced affinities, these fragments must be screened using more
sensitive biophysical assays, such as Fluorescence-Based Thermal Shift, NMR Spectroscopy and
Surface Plasmon Resonance [142]. Molecular docking can also be an invaluable tool for the detection
of potentially interacting fragments and several examples will be discussed below. Candidate
fragments detected by experimental or computational approaches are then usually evaluated
through X-ray Crystallography [142] or even High Throughput X-ray Crystallography (HTX), where
protein crystals are soaked in high concentrations of one or more fragments and the structure of the
complex is subsequently determined [143].

4.2. Fragment Libraries

Some aspects must be taken into consideration when tailoring fragment libraries in order to
optimise fragment-based drug design (FBDD) outcomes. For instance, because fragments are smaller,
they tend to bind less tightly to the protein targets, exhibiting lower potency values. Therefore, it is
advantageous to use size-normalised parameters, such as Ligand Efficiency (LE) [144], Binding
Efficiency Index (BEI) [145] or Fit Quality (FQ) [146], to prioritise the evaluated molecules. These can
then serve as objective parameters to a successful subsequent lead optimisation [147]. Secondly,
Harren Jhoti’s group has suggested an adjusted set of rules [148] (or guidelines [149]), termed Rule
of Three (RO3), derived from hits obtained via High Throughput X-ray Crystallography (HTX) and
inspired by Lipinski’s Rule of Five [150]. These stemmed from the observation that successful hits
customarily present molecular weight under 300 Da, three or fewer hydrogen bond donors, three or
Int. J. Mol. Sci. 2019, 20, 4574 14 of 29

fewer hydrogen bond acceptors, clogP under three and, additionally, three or fewer rotatable bonds
and a polar surface area under 60 Å2. These guidelines can help filtering fragment libraries for
efficient screening, both experimentally and computationally. A third matter worth noting is the
reported “lack of tri-dimensionality” in fragment libraries, which can hinder the development of
ligands with high affinity for certain classes of targets [151].
Fragment libraries can be generic or generated ad hoc (targeted, or focused libraries). Many of
the generic libraries are commercially available on demand, and thus may be readily used in
experimental screens, and the compound chemical structures are usually also available as Structure-
Data Files (SDF), which can be straightforwardly converted to other structural formats, such as MOL2,
PDB and PDBQT and used for virtual screening (cf. Verheij [152] work on lead-likeness for sources
of such libraries). Fragment libraries usually contain 102 to 104 molecules, which are generally
compliant with the RO3 and are idealised to maximise attributes such as solubility, chemical stability,
scaffold complexity, tri-dimensionality and tractability [151,153,154]. Tractability-guided
fragmentation algorithms and pipelines can be used to generate specialised fragment libraries
starting from collections such as the World Drug Index (which has been fragmented using the RECAP
algorithm [155]) or natural products libraries [156].
The combination of fragments into a larger molecule has been classified into four distinct
categories, namely Merging, Linking, Growing and “SAR by catalogue” [153]. In fragment merging,
two fragments occupying an overlapping site are joined together to obtain a larger molecule with
higher affinity. Conversely, in fragment linking, the fragments are usually bound to two distinct
binding pockets (or sub-pockets) and are joined together via the construction of a linker fragment,
that ideally allows the maintenance of the initial orientation of the fragments. Fragment growing
consists of the design and incorporation of new functional groups that are expected to form new
interactions with the receptor, thus increasing the binding affinity. Finally, the “SAR by catalogue”
is particularly interesting from the virtual screening angle due to its simplicity; in this approach, a
fragment initially detected (and ideally confirmed by experimental techniques) is then used as an
“anchor” to query a database for larger molecules that contain the original fragment. Thus, effectively,
this strategy is largely used to create more focused libraries.

4.3. Molecular Docking in FBDD

Many groups have used FBDD to idealise potent ligands for disease-modifying protein targets
with extensive use of molecular docking and virtual screening approaches. In a study developed by
Chen and Shoichet, a fragment-based approach was used as an alternative to a lead-like virtual screen
campaign, obtaining increased hit rates for β-Lactamase inhibitors, and ultimately yielding hits in the
low μM range [157]. This indicates that even using similar docking protocols, fragment-based
approaches can yield more accurate initial hits when compared to lead-like molecules screening.
These computationally-driven works reflect some of the experimental strategies discussed
above, since the initial screens for promising fragments are usually followed by a fragment-joining
step, which can be accomplished in a manual [158] or automated [159] way. Recently, Park et al. have
been able to design nanomolar-range inhibitors for the protein Glycogen Synthase Kinase-3 β, using
AutoDock [32] as the initial tool to perform virtual screening of fragment libraries in three
independent subsites and LigBuilder [160] as the tool to connect a series of selected fragments [159].
Employing the “SAR by catalogue” method, Zhao and co-workers, after initial filtering of the
ZINC database, have used an in-house docking solution to prioritise anchor fragments that bind the
BRD4 bromodomain, which were then used to further interrogate the database and retrieve
compounds containing the selected moieties, ultimately finding compounds with activity in the low
micromolar range (7.0–7.5 μM) [161]. Using a similar “anchor-based” analogue search approach,
Rudling and co-workers have used Dock3.6 to find inhibitors in the low micromolar range for MTH1
protein, an interesting cancer target, and in a second round of prospection for commercially available
analogues, they managed to further optimise the initial hits to achieve IC50 values as low as 9 nM
[162].
Int. J. Mol. Sci. 2019, 20, 4574 15 of 29

Hernandez et al. have suggested non-nucleoside inhibitors of flaviviral methyltransferases (Zika

virus and Dengue virus NS5MTase) presenting IC50 ~20 μM by screening a focused library
constructed using a knowingly binding core substructure, encoded organic chemistry rules and
commercially available building blocks. The authors refer to this approach as fragment-growing
[163].
The successful combination of fragment-based virtual screening and NMR screening has also
been reported. Fjellström et al. have identified Activated Factor XI inhibitors using Glide to prioritise
1800 molecules (out of 6.5 × 103 from AstraZeneca screening collection with molecular weight (MW)
< 250 g/mol) for NMR fragment screening. Subsequent structure-based expansion and re-scoring of
13 NMR hits yielded a compound with activity of 1.0 nM [158]. Using an inverted approach,
Akabayov and co-workers used an initial NMR screen of a library containing 1000 fragments to
identify moieties that bind T7 DNA primase, the two most promising hits were then used to query
ZINC database, once more reflecting the “SAR by catalogue” approach, and the selected molecules
(approximately 3000 per scaffold) were docked to DNA primase structure, using Autodock4. About
half of the 16 selected compounds showed inhibitory activities [164].
Amaning and co-workers prospected for MEK1 inhibitors carrying out a virtual screening
campaign of approximately 104 molecules, used to prioritise fragments to be further characterised by
differential scanning fluorimetry (DSF), surface plasmon resonance and X-ray crystallography.
Interestingly, a parallel biochemical screening of the same library showed that the 5% of the best
scoring molecules in the virtual screening contained 30% of the biochemical hits and, according to
the authors, this indicates that the VS–DSF combination can used to ‘jump-start’ a project in an early
phase when a biochemical or other biophysical assays are not available [165]. Additionally, it has
been suggested that characteristics such as novelty and potency are likely to differ considerably
between hits determined by experimental screening and those determined by virtual screening [166].
Besides prospecting for new molecules in fragment-based VS campaigns, molecular docking is
extensively used to hypothesise interaction modes and better characterise the ligand-receptor
interactions [167–169], and remains an invaluable asset in the drug development toolkit.

5. Machine Learning-Based Approaches

Scoring and ranking candidate molecules through binding affinity prediction is the most
challenging aspect of molecular docking and VS. Classical SFs must simplify and generalise many
aspects of the receptor-ligand interaction in order to maintain efficiency, approachability and
accessibility [27]. Moreover, these SFs employ linear regression models: parametric supervised
learning methods, which assume a specific predetermined functional form [170]. In other words,
parametric methods fit the input variables (such as van der Waals and electrostatic energy terms) to
the output (binding energy score) into a function whose form is already specified, and which is
adjusted during the development of the SF in a theory-inspired fashion [77]. This rigid scheme often
results in unadaptable SFs which fail to capture intrinsic nonlinearities in the data and therefore
underperform in situations not accounted for in their formulation [77,171].
Alternatively, nonparametric machine learning (ML) algorithms (often referred to as just
“machine learning”) can be used to replace [77,172–174] or improve [82,175–178] predetermined
functional forms in classical SFs for binding affinity predictions. They have also been successfully
applied in binders/nonbinders identification in virtual screening [175,179–181] and native pose
prediction [126,172,182].
ML methods are divided into two broad groups: supervised and unsupervised learning.
Unsupervised learning algorithms are employed to model the training data when there is no output
available. Thus, these algorithms are commonly used for clustering data based on the degree of
similarity between their features, for detecting associations between the data points, and for density
estimations. In supervised learning, however, the output variables are known and provided to the
algorithm along with the input for training. In nonparametric supervised learning, no functional form
is assumed. It is then possible to infer the correlations between input and output from the training
data itself and utilise it to predict the output for datasets of which the outcomes are unknown [170].
Int. J. Mol. Sci. 2019, 20, 4574 16 of 29

This allows for more diverse and accurate SFs: more features from the docked complex can be
accounted for implicitly, therefore skirting modelling assumptions and necessary generalisations of
classical SFs [77,82,171]. Moreover, the adeptness of the ML algorithm can be adjusted by tailoring
the training dataset. For instance, increasing the diversity of the training complexes results in ML SFs
with greater comprehensiveness. In fact, it has been shown that increasing the size of the training set
boosts the scoring function’s performance [82,172,183].
This contrasts greatly with classical SFs, whose parametric behaviour remains unable to improve
performance with larger training datasets [82]. On the other hand, increasing the level of feature
detail in training sets comprising of similar complexes may provide greater discrimination power
when studying such data [183,184].

5.1. Protein Target Types: Generic and Family-Specific

Machine learning SFs can be considered family-specific or generic. It has been shown that
family-specific SFs can outperform most accurate generic ones at said protein family’s predictions
[183,184]. Until recently, however, it was not clear whether a family-specific SF carried any
advantages over generic ones whose training includes all complexes and features utilised in training
the former [83]. It was later shown that random forest trained with family-specific data only slightly
outperformed the universal model. This outperformance grew, however, when predicting more
difficult targets with less active ligands [185]. In a 2018 study with deep learning neural networks,
Imrie et al. [183] showed that family-specific models trained with a subset of the entire dataset
outperformed universally trained models, and that only limited family data was required for this
outperformance to occur. For each different protein family, the importance of the features used to
describe the data varies [184], therefore, specific SFs are able to better assimilate these characteristics
as a result of dealing with less broad and more nuanced data [183–185].
Machine learning SFs have been regarded both as knowledge-based [186,187] and empirical
[188]. However, it is important to note that this categorisation has extensively been used in regard to
classical SFs, and therefore it should not obscure the fact that there is a more fundamental difference
between them: the former consists of nonparametric and the latter of parametric learning (Figure 4).

Figure 4. Learning methods can be broadly divided into supervised learning, when there is data
available for training and parameterisation; and unsupervised learning, when there is no such data.
Unsupervised learning cannot be used for binding affinity predictions and virtual screening.
Int. J. Mol. Sci. 2019, 20, 4574 17 of 29

Supervised learning, on the other hand, can be divided into parametric and nonparametric learning.
Parametric learning assumes a predetermined functional form, as observed in linear regression, and
is the method employed in classical scoring functions. Nonparametric learning, or just machine
learning, does not presume a predetermined functional form, which is instead inferred from the data
itself. It can yield continuous output, as in nonlinear regression, or discrete output, for classification
problems such as binders/nonbinders identification.

5.2. Experiment Types: Binding Affinity Prediction and Virtual Screening

SFs designed for binding affinity predictions can also be used for virtual screening experiments,
as long as the predicted results are ordered from best to worst binding score. If a binary
active/inactive distinction is desired, one can establish an optimal activity threshold score by
analysing the SF’s performance on a benchmarking dataset (c.f. Benchmark Datasets section).
However, ML classifiers built for VS may present better discrimination since their training utilises
datasets specific for portraying virtual screening circumstances i.e., they are often trained on data
derived from in silico approaches (as opposed to crystal structures of complexes) which do not
always represent the correct binding mode, and the features from docked decoy molecules are also
used for training [189].

5.3. Algorithms and Feature Selection

Feature selection plays an important role in the development of ML methods. Selecting a subset
of features which are appropriate and effective for characterising the data not only improves
prediction performance, but also reduces computational expense and facilitates the understanding of
the intrinsic patterns underlying the data [190].
The first ML SF to outperform classical SFs [83], random forest (RF)-Score [77], utilised the
random forest (RF) algorithm with intermolecular interaction features comprised of the number of a
particular protein-ligand atom type pair interacting within a certain distance range [77]. Other
descriptors such as energy terms from classical SFs, solvent accessible surface area, entropy,
hydrophobic interactions and chemical descriptors have been applied by works such as those of
Springer et al. (PostDOCK) [181], Pereira et al. (DeepVS) [177], Jiménez et al. (Kdeep) [78], Durrant et
al. (NNScore) [79], Koppisetty et al. [191] and Liu et al. (B2BScore) [192] with various degrees of
success. It has been shown that richer and more precise chemical descriptors do not generally result
in more accurate predictions [193], and that different SFs have very different responses to an increase
in the number of features [171].
Other ways of describing the data have been explored. For instance, Kundu et al. [194] utilised
fundamental molecular descriptors for the proteins and the ligands, without any intermolecular
interaction features, which circumvents the need for binding pose information. Srinivas et al. [195]
utilised collaborative filtering, an algorithm extensively employed for recommendation systems (i.e.,
predicting appropriate online costumer recommendations), to bypass the explicit definition of
receptor and ligand features. The similarities in the data are inferred only based on the results of the
recorded binding assays.

5.4. Deep Learning

Deep learning neural networks have recently been applied to pose prediction and ranking
[78,173,177,183,196]. Convolutional neural networks, which are known to present outstanding image
recognition capabilities [197], in molecular docking, have been explored mainly by featurising the
protein-ligand complexes as three-dimensional grids. Deep learning SFs have yielded state-of-the-art
results [78,183,196], comparable to and even surpassing those achieved by random forest, support
vector machines, and boosted regression trees, the other non-neural network algorithms reported to
be the most accurate for protein-ligand scoring [171,198].

5.5. Recent Applications and Perspectives

Int. J. Mol. Sci. 2019, 20, 4574 18 of 29

It is noteworthy that although the current ML techniques already promise to advance

computational drug discovery, some limitations still need to be addressed. For instance, larger
amounts of data are still required to reach optimal deep learning performance, and it is not clear
whether at some point learning saturation can occur [183]. Furthermore, complex nonparametric
learning models can be difficult to interpret. Sieg et al. [199] very recently pointed out that bias is
being implicitly learned from standard benchmarking sets, and suggested guidelines to avoid
fallacious models.
ML SFs for molecular docking have only recently been introduced. Naturally, most studies are
dedicated to assessing and improving their predictive powers, and not as many have applied them
in drug discovery and repurposing experiments. Nonetheless, existing prospective studies show
positive results (Table 4). In 2011, Kinnings et al. [175] created a support vector machine-based SF to
improve binding affinity prediction from classical SFs and used it to identify that phosphodiesterase
inhibitors could potentially be repurposed towards Mycobacterium tuberculosis protein InhA. One year
later, Zhan et al. [123] used support vector machine to integrate classical docking scores, interaction
profiles and molecular descriptors to identify six novel Akt1 inhibitors. Durrant et al. (2015) used
NNScore, a neural network SF, to describe 39 novel oestrogen-receptor ligands, whose activities were
experimentally confirmed [200].
Among the ML SFs mentioned in this section, those readily accessible for use are the following:
RF-Score; NN-Score; Ragoza et al.’s final optimised model architecture; DLScore; and kDEEP. These
are available as downloadable standalone programs, with the exception of Kdeep, which can be
found at playmolecule.org. If online docking is desired, CSM-lig [201] (for binding affinity
predictions) is also available as a web-server. To the best of our knowledge, none of these SFs have
been integrated into docking programs such as the ones summarised in Table 4.
Machine learning methods have shown positive results, as well as promising room for more
enhancement. In addition, the availability of benchmarking data for training and testing is likely to
be further expanded, which will consequently improve the predictive power of these techniques.
Therefore, nonparametric machine learning is potentially the next step to drastically improve
molecular docking predictiveness and accuracy.

Table 4. Recent developments using machine learning (ML) algorithms in molecular docking.

Generic or Type of
ML Training
SF Name Best Performance Family Docking Reference
Algorithm Database
Specific Study
Ballester
RF-Score RFa PDBbind Rpb = 0.776 Generic BAPc
2010 [77]
Liu 2013
B2BScore RF PDBbind Rp = 0.746 Generic BAP
[192]
Zilian, 2013
SFCScoreRF RF PDBbind Rp = 0.779 Generic BAP
[202]
Constructed Springer,
PostDOCK RF 92% accuracy Generic VSd
from PDB 2005 [181]
Kinnings,
- SVMe DUD - Both VS
2011 [175]
Li, 2013
ID-Score SVRf PDBbind Rp = 0.85 Generic BAP
[203]
PDB; MOAD; Durrant,
NNScore NNg EF = 10.3 Generic VS
PDBbind-CN 2010 [79]
Rp = 0.7668 (gen.)
Ouyang,
CScore NN PDBbind Rp = 0.8237 (fam. Both BAP
2011 [174]
spec.)
Ragoza,
- Deep NN CSAR, DUD-E ROCAUC = 0.868 Generic VS
2017 [196]
Imrie, 2018
- Deep NN DUD-E ROCAUC = 0.92 Both VS
[183]
Int. J. Mol. Sci. 2019, 20, 4574 19 of 29

Hassan,
DLScore Deep NN PDBbind Rp = 0.82 Generic BAP
2018 [173]
Pereira,
DeepVS Deep NN DUD ROCAUC = 0.81 Generic VS
2016 [177]
1. Jiméne
Kdeep Deep NN PDBbind Rp = 0.82 Generic BAP z, 2018
[78]
a
Random Forest; bPearson’s Correlation Coefficient; cBinding Affinity Prediction; dVirtual Screening;
e
Support Vector Machine; fSupport Vector Regression; gNeural Network.

5. Conclusions
Molecular docking has been established as a pivotal technique among the computational tools
for structure-based drug discovery. Here we addressed key aspects of the methodology and
discussed recent trends in the literature for advancing and employing the technique for successful
drug design. Benchmarking sets and the various metrics available are crucial for validating
performance gains achieved by new docking software but must be carefully chosen since no single
one can be regarded as the absolute best for molecular docking. A significant improvement in the
performance of all docking software can be achieved by employing multiple SFs for consensus posing
and/or scoring. As reviewed here, there is a plethora of protocols for consensus docking to be
explored by the user.
FBDD emerged as a successful paradigm for developing new drugs, combining the serendipity
of target-based high throughput screening with the rationality of structure-based drug design
approaches. Molecular docking has important roles in FBDD, from planning and prioritisation of
fragment library composition to finding analogues with improved binding affinities through large-
scale VS of compound libraries.
ML is a branch of artificial intelligence that has gained much attention in diverse fields of science
and technology and molecular docking methods are also taking advantage of this pulsating area.
Although recent, the flexibility of ML in modelling data has already rendered more diverse and
accurate SFs implicitly accounting for more features from the docked complex.

Funding: FPSJr is a productivity fellow from the National Council of Technological and Scientific Development
(CNPq) and holds a Newton Advanced Fellowship from the United Kingdom Academy of Medical Sciences.
PHMT research is funded by The Cystic Fibrosis Trust (SRC 010 - RG92232). PJ is a M.Sc. grantee supported by
the “Coordenação de Aperfeiçoamento de Pessoal de Nível Superior” (CAPES - Brazil).

Acknowledgments: The authors thank the Brazilian National Council for Research and Development (CNPq),
the State of Rio de Janeiro Research Foundation (FAPERJ), the Oswaldo Cruz Foundation for general financial
support. The authors also thank the researchers from the Blundell Group from the University of Cambridge for
helpful discussions and Dr. Isabella Alvim Guedes from the National Laboratory for Scientific Computation
(LNCC, Petrópolis, Brazil) for insightful clarifications regarding DockThor’s algorithm.

Conflicts of Interest: The authors declare no conflict of interest. The funders had no role in the design of the
study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to
publish the results.

Abbreviations
ML Machine Learning
RF Random Forest
MD Molecular Dynamics
SF Scoring Function
FBDD Fragment-Based Drug Design
VS Virtual Screening
MW Molecular Weight
SAR Structure-Activity Relationship
QSAR Quantitative Structure-Activity Relationship
Int. J. Mol. Sci. 2019, 20, 4574 20 of 29

EF Enrichment Factor
ROC Receiver Operating Characteristic
PDB Protein Data Bank
RMSD Root-Mean-Square Deviation
MUV Maximum Unbiased Validation
DUD Directory of Useful Decoys
GPCR G-Protein-Coupled Receptor
LADS Latent Actives in the Decoy Set
BEDROC Boltzmann-Enhanced Discrimination of Receiver Operating Characteristic
AUC Area Under the Curve
RIE Robust Initial Enhancement
DUD-E Directory of Useful Decoys, Enhanced
DEKOIS Demanding Evaluation Kits for Objective in Silico Screening
HTX High Throughput X-ray Crystallography
RO3 Rule of Three
DSF Differential Scanning Fluorimetry
Rp Pearson correlation coefficient
Rs Spearman rank-correlation
BFGS Broyden–Fletcher–Goldfarb–Shanno

References
1. Liu, Y.; Zhang, Y.; Zhong, H.; Jiang, Y.; Li, Z.; Zeng, G.; Chen, M.; Shao, B.; Liu, Z.; Liu, Y. Application of
molecular docking for the degradation of organic pollutants in the environmental remediation: A review.
Chemosphere 2018, 203, 139–150.
2. Morris, G.M.; Goodsell, D.S.; Halliday, R.S.; Huey, R.; Hart, W.E.; Belew, R.K.; Olson, A.J. Automated
docking using a Lamarckian genetic algorithm and an empirical binding free energy function. J. Comput.
Chem. 1998, 19, 1639–1662.
3. Trott, O.; Olson, A.J. AutoDock Vina: Improving the speed and accuracy of docking with a new scoring
function, efficient optimization, and multithreading. J. Comput. Chem. 2009, 28, 455–461.
4. De Magalhães, C.S.; Almeida, D.M.; Barbosa, H.J.C.; Dardenne, L.E. A dynamic niching genetic algorithm
strategy for docking highly flexible ligands. Inf. Sci. (Ny) 2014, 289, 206–224.
5. de Magalhães, C.S.; Barbosa, H.J.C.; Dardenne, L.E. Selection-Insertion Schemes in Genetic Algorithms for
the Flexible Ligand Docking Problem. Lect. Notes Comput. Sci. 2004, 3102, 368–379.
6. Jones, G.; Willett, P.; Glen, R.C.; Leach, A.R.; Taylor, R.; Uk, K.B.R. Development and Validation of a Genetic
Algorithm for Flexible Docking. J. Mol. Biol. 1997, 267, 727–748.
7. Verdonk, M.L.; Cole, J.C.; Hartshorn, M.J.; Murray, C.W.; Taylor, R.D. Improved protein-ligand docking
using GOLD. Proteins Struct. Funct. Genet. 2003, 52, 609–623.
8. Rarey, M.; Kramer, B.; Lengauer, T.; Klebe, G. A fast flexible docking method using an incremental
construction algorithm. J. Mol. Biol. 1996, 261, 470–489.
9. Thomsen, R.; Christensen, M.H. MolDock: A new technique for high-accuracy molecular docking. J. Med.
Chem. 2006, 49, 3315–3321.
10. Gioia, D.; Bertazzo, M.; Recanatini, M.; Masetti, M.; Cavalli, A. Dynamic docking: A paradigm shift in
computational drug discovery. Molecules 2017, 22, 2029.
11. Berman, H.M. The Protein Data Bank. Nucleic Acids Res. 2000, 28, 235–242.
12. Hetényi, C.; Van Der Spoel, D. Blind docking of drug-sized compounds to proteins with up to a thousand
residues. FEBS Lett. 2006, 580, 1447–1450.
13. Volkamer, A.; Kuhn, D.; Grombacher, T.; Rippmann, F.; Rarey, M. Combining global and local measures
for structure-based druggability predictions. J. Chem. Inf. Model. 2012, 52, 360–372.
14. Radoux, C.J.; Olsson, T.S.G.; Pitt, W.R.; Groom, C.R.; Blundell, T.L. Identifying Interactions that Determine
Fragment Binding at Protein Hotspots. J. Med. Chem. 2016, 59, 4314–4325.
15. Fu, D.Y.; Meiler, J. Predictive Power of Different Types of Experimental Restraints in Small Molecule
Docking: A Review. J. Chem. Inf. Model. 2018, 58, 225–233.
16. Brooijmans, N.; Kuntz, I.D. Molecular Recognition and Docking Algorithms. Annu. Rev. Biophys. Biomol.
Struct. 2003, 32, 335–373.
Int. J. Mol. Sci. 2019, 20, 4574 21 of 29

17. Meng, E.C.; Shoichet, B.K.; Kuntz, I.D. Automated docking with grid-based energy evaluation. J. Comput.
Chem. 1992, 13, 505–524.
18. Irwin, J.J.; Shoichet, B.K. ZINC—A Free Database of Commercially Available Compounds for Virtual
Screening. J. Chem. Inf. Model. 2006, 45, 177–182.
19. Kim, S.; Thiessen, P.A.; Bolton, E.E.; Chen, J.; Fu, G.; Gindulyte, A.; Han, L.; He, J.; He, S.; Shoemaker, B.A.;
et al. PubChem Substance and Compound databases. Nucleic Acids Res. 2016, 44, D1202–D1213.
20. Hanwell, M.D.; Curtis, D.E.; Lonie, D.C.; Vandermeerschd, T.; Zurek, E.; Hutchison, G.R. Avogadro: An
advanced semantic chemical editor, visualization, and analysis platform. J. Cheminform. 2012, 4, 17.
21. Pearlman, R.S. Rapid Generation of High Quality Approximate 3-dimension Molecular Structures. Chem.
Des. Auto. News 1987, 2, 1–7.
22. McCammon, J.A.; Nielsen, J.E.; Baker, N.A.; Dolinsky, T.J. PDB2PQR: An automated pipeline for the setup
of Poisson–Boltzmann electrostatics calculations. Nucleic Acids Res. 2004, 32, W665–W667.
23. Anandakrishnan, R.; Aguilar, B.; Onufriev, A.V. H++ 3.0: Automating pK prediction and the preparation
of biomolecular structures for atomistic molecular modeling and simulations. Nucleic Acids Res. 2012, 40,
W537–W541.
24. Forli, S.; Huey, R.; Pique, M.E.; Sanner, M.F.; Goodsell, D.S.; Olson, A.J. Computational protein-ligand
docking and virtual drug screening with the AutoDock suite. Nat. Protoc. 2016, 11, 905–919.
25. Dardenne, L.E.; Barbosa, H.J.C.; De Magalhães, C.S.; Almeida, D.M.; da Silva, E.K.; Custódio, F.L.; Guedes,
I.A. DockThor Portal. Available online: https://dockthor.lncc.br/v2/ (accessed on 22 March 2019).
26. Guedes, I.A.; de Magalhães, C.S.; Dardenne, L.E. Receptor-ligand molecular docking. Biophys. Rev. 2014, 6,
75–87.
27. Kitchen, D.B.; Decornez, H.; Furr, J.R.; Bajorath, J. DOCKING and scoring in virtual screening for drug
discovery: Methods and applications. Nat. Rev. Drug Discov. 2004, 3, 935.
28. Zsoldos, Z.; Reid, D.; Simon, A.; Sadjad, B.S.; Johnson, A.P. eHiTS: An Innovative Approach to the Docking
and Scoring Function Problems. Curr. Protein Pept. Sci. 2006, 7, 421–435.
29. Moitessier, N.; Englebienne, P.; Lee, D.; Lawandi, J.; Corbeil, C.R. Towards the development of universal,
fast and highly accurate docking/scoring methods: A long way to go. Br. J. Pharmacol. 2008, 153, 7–26.
30. Hindle, S.A.; Rarey, M.; Buning, C.; Lengauer, T. Flexible docking under pharmacophore type constraints.
J. Comput. Aided Mol. Des. 2002, 16, 129–149.
31. Huey, R.; Morris, G.M.; Olson, A.J.; Goodsell, D.S. A semiempirical free energy force field with charge-
based desolvation. J. Comput. Chem. 2007, 28, 1145–1152.
32. Morris, G.M.; Huey, R.; Lindstrom, W.; Sanner, M.F.; Belew, R.K.; Goodsell, D.S.; Olson, A.J. AutoDock4
and AutoDockTools4: Automated docking with selective receptor flexibility. J. Comput. Chem. 2009, 30,
2785–2791.
33. Koes, D.R.; Baumgartner, M.P.; Camacho, C.J. Lessons learned in empirical scoring with smina from the
CSAR 2011 benchmarking exercise. J. Chem. Inf. Model. 2013, 53, 1893–1904.
34. Korb, O.; Stützle, T.; Exner, T.E. Empirical scoring functions for advanced Protein-Ligand docking with
PLANTS. J. Chem. Inf. Model. 2009, 49, 84–96.
35. Korb, O.; Stützle, T.; Exner, T.E. An ant colony optimization approach to flexible protein–ligand docking.
Swarm Intell. 2007, 1, 115–134.
36. Abagyan, R.; Kuznetsov, D.; Totrov, M. ICM-New Method for Protein Modeling and Design: Applications
to Docking and Structure Prediction from. J. Comput. Chem. 1994, 15, 488–506.
37. Abagyan, R.; Totrov, M. Biased probability Monte Carlo conformational searches and electrostatic
calculations for peptides and proteins. J. Mol. Biol. 1994, 235, 983–1002.
38. Friesner, R.A.; Banks, J.L.; Murphy, R.B.; Halgren, T.A.; Klicic, J.J.; Mainz, D.T.; Repasky, M.P.; Knoll, E.H.;
Shelley, M.; Perry, J.K.; et al. Glide: A new approach for rapid, accurate docking and scoring. 1. Method
and assessment of docking accuracy. J. Med. Chem. 2004, 47, 1739–1749.
39. Jain, A.N. Surflex: Fully automatic flexible molecular docking using a molecular similarity-based search
engine. J. Med. Chem. 2003, 46, 499–511.
40. Jain, A.N. Surflex-Dock 2.1: Robust performance from ligand energetic modeling, ring flexibility, and
knowledge-based search. J. Comput. Aided Mol. Des. 2007, 21, 281–306.
41. Yang, J.M.; Chen, C.C. GEMDOCK: A Generic Evolutionary Method for Molecular Docking. Proteins Struct.
Funct. Bioinform. 2004, 55, 288–304.
Int. J. Mol. Sci. 2019, 20, 4574 22 of 29

42. Allen, W.J.; Balius, T.E.; Mukherjee, S.; Brozell, S.R.; Moustakas, D.T.; Lang, P.T.; Case, D.A.; Kuntz, I.D.;
Rizzo, R.C. DOCK 6: Impact of new features and current docking performance. J. Comput. Chem. 2015, 36,
1132–1156.
43. Li, H.; Li, C.; Gui, C.; Luo, X.; Chen, K.; Shen, J.; Wang, X.; Jiang, H. GAsDock: A new approach for rapid
flexible docking based on an improved multi-population genetic algorithm. Bioorg. Med. Chem. Lett. 2004,
14, 4671–4676.
44. Rarey, M.; Wefing, S.; Lengauer, T. Placement of medium-sized molecular fragments into active sites of
proteins. J. Comput. Aided Mol. Des. 1996, 10, 41–54.
45. McGann, M. FRED pose prediction and virtual screening accuracy. J. Chem. Inf. Model. 2011, 51, 578–596.
46. Plewczynski, D.; Łaźniewski, M.; Augustyniak, R.; Ginalski, K. Can we trust docking results? Evaluation
of seven commonly used programs on PDBbind database. J. Comput. Chem. 2011, 32, 742–755.
47. Chang, M.W.; Ayeni, C.; Breuer, S.; Torbett, B.E. Virtual screening for HIV protease inhibitors: A
comparison of AutoDock 4 and Vina. PLoS ONE 2010, 5, e11955.
48. Capoferri, L.; Leth, R.; ter Haar, E.; Mohanty, A.K.; Grootenhuis, P.D.J.; Vottero, E.; Commandeur, J.N.M.;
Vermeulen, N.P.E.; Jørgensen, F.S.; Olsen, L.; et al. Insights into regioselective metabolism of mefenamic
acid by cytochrome P450 BM3 mutants through crystallography, docking, molecular dynamics, and free
energy calculations. Proteins Struct. Funct. Bioinform. 2016, 84, 383–396.
49. Feng, Z.; Pearce, L.V.; Xu, X.; Yang, X.; Yang, P.; Blumberg, P.M.; Xie, X.-Q. Structural Insight into
Tetrameric hTRPV1 from Homology Modeling, Molecular Docking, Molecular Dynamics Simulation,
Virtual Screening, and Bioassay Validations. J. Chem. Inf. Model. 2015, 55, 572–588.
50. Vadloori, B.; Sharath, A.K.; Prabhu, N.P.; Maurya, R. Homology modelling, molecular docking, and
molecular dynamics simulations reveal the inhibition of Leishmania donovani dihydrofolate reductase-
thymidylate synthase enzyme by Withaferin-A. BMC Res. Notes 2018, 11, 246.
51. Yadav, D.K.; Kumar, S.; Misra, S.; Yadav, L.; Teli, M.; Sharma, P.; Chaudhary, S.; Kumar, N.; Choi, E.H.;
Kim, H.S.; et al. Molecular Insights into the Interaction of RONS and Thieno [3, 2-c]pyran Analogs with
SIRT6/COX-2: A Molecular Dynamics Study. Sci. Rep. 2018, 8, 4777.
52. Makhouri, F.R.; Ghasemi, J.B. Combating Diseases with Computational Strategies Used for Drug Design
and Discovery. Curr. Top. Med. Chem. 2019, 18, 2743–2773.
53. Wang, Z.; Sun, H.; Yao, X.; Li, D.; Xu, L.; Li, Y.; Tian, S.; Hou, T. Comprehensive evaluation of ten docking
programs on a diverse set of protein-ligand complexes: The prediction accuracy of sampling power and
scoring power. Phys. Chem. Chem. Phys. 2016, 18, 12964–12975.
54. Halgren, T.A. Merck molecular force field. I. Basis, form, scope, parameterization, and performance of
MMFF94. J. Comput. Chem. 1996, 17, 490–519.
55. Hingerty, B.E.; Ritchie, R.H.; Ferrell, T.L.; Turner, J.E. Dielectric effects in biopolymers: The theory of ionic
saturation revisited. Biopolymers 1985, 24, 427–439.
56. Halgren, T.A. The representation of van der Waals (vdW) interactions in molecular mechanics force fields:
Potential form, combination rules, and vdW parameters. J. Am. Chem. Soc. 1992, 114, 7827–7843.
57. Hansch, C.; Fujita, T. ρ-σ-π Analysis. A Method for the Correlation of Biological Activity and Chemical
Structure. J. Am. Chem. Soc. 1964, 86, 1616–1626.
58. 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. 1997, 11, 425–445.
59. Friesner, R.A.; Murphy, R.B.; Repasky, M.P.; Frye, L.L.; Greenwood, J.R.; Halgren, T.A.; Sanschagrin, P.C.;
Mainz, D.T. Extra precision glide: Docking and scoring incorporating a model of hydrophobic enclosure
for protein-ligand complexes. J. Med. Chem. 2006, 49, 6177–6196.
60. Velec, H.F.G.; Gohlke, H.; Klebe, G. DrugScore(CSD)-knowledge-based scoring function derived from
small molecule crystal data with superior recognition rate of near-native ligand poses and better affinity
prediction. J. Med. Chem. 2005, 48, 6296–6303.
61. Muegge, I. PMF scoring revisited. J. Med. Chem. 2006, 49, 5895–5902.
62. Bohacek, R.S.; McMartin, C.; Guida, W.C. The art and practice of structure-based drug design: A molecular
modeling perspective. Med. Res. Rev. 1996, 16, 3–50.
63. Amaro, R.E.; Baudry, J.; Chodera, J.; Demir, Ö.; McCammon, J.A.; Miao, Y.; Smith, J.C. Ensemble Docking
in Drug Discovery. Biophys. J. 2018, 114, 2271–2278.
Int. J. Mol. Sci. 2019, 20, 4574 23 of 29

64. Korb, O.; Olsson, T.S.G.; Bowden, S.J.; Hall, R.J.; Verdonk, M.L.; Liebeschuetz, J.W.; Cole, J.C. Potential and
limitations of ensemble docking. J. Chem. Inf. Model. 2012, 52, 1262–1274.
65. Totrov, M.; Abagyan, R. Flexible ligand docking to multiple receptor conformations: A practical alternative.
Curr. Opin. Struct. Biol. 2008, 18, 178–184.
66. De Paris, R.; Vahl Quevedo, C.; Ruiz, D.D.; Gargano, F.; de Souza, O.N. A selective method for optimizing
ensemble docking-based experiments on an InhA Fully-Flexible receptor model. BMC Bioinform. 2018, 19,
235.
67. De Paris, R.; Frantz, F.A.; Norberto de Souza, O.; Ruiz, D.D.A. wFReDoW: A Cloud-Based Web
Environment to Handle Molecular Docking Simulations of a Fully Flexible Receptor Model. BioMed Res.
Int. 2013, 2013, 469363.
68. Cavasotto, C.N.; Kovacs, J.A.; Abagyan, R.A. Representing receptor flexibility in ligand docking through
relevant normal modes. J. Am. Chem. Soc. 2005, 127, 9632–9640.
69. Damm, K.L.; Carlson, H.A. Exploring experimental sources of multiple protein conformations in structure-
based drug design. J. Am. Chem. Soc. 2007, 129, 8225–8235.
70. Leach, A.R.; Shoichet, B.K.; Peishoff, C.E. Prediction of protein-ligand interactions. Docking and scoring:
successes and gaps. J. Med. Chem. 2006, 49, 5851–5855
71. Wang, R.; Fang, X.; Lu, Y. The PDBbind Database:  Collection of Binding Affinities for Protein−Ligand
Complexes with Known Three-Dimensional Structures—Journal of Medicinal Chemistry (ACS
Publications). J. Med. Chem. 2004, 47, 2977–2980.
72. Ahmed, A.; Smith, R.D.; Clark, J.J.; Dunbar, J.B., Jr.; Carlson, H.A. Recent improvements to Binding MOAD:
A resource for protein-ligand Binding affinities and structures. Nucleic Acids Res. 2015, 43, D465–D469.
73. Smith, R.D.; Ung, P.M.-U.; Esposito, E.X.; Wang, S.; Carlson, H.A.; Dunbar, J.B.; Yang, C.-Y. CSAR
Benchmark Exercise of 2010: Combined Evaluation Across All Submitted Scoring Functions. J. Chem. Inf.
Model. 2011, 51, 2115–2131.
74. Block, P. AffinDB: A freely accessible database of affinities for protein-ligand complexes from the PDB.
Nucleic Acids Res. 2006, 34, D522–D526.
75. Wang, R.; Fang, X.; Lu, Y.; Yang, C.Y.; Wang, S. The PDBbind database: Methodologies and updates. J. Med.
Chem. 2005, 48, 4111–4119.
76. Zhao, Z.; Liu, J.; Wang, R.; Liu, Z.; Liu, Y.; Han, L.; Li, Y.; Nie, W.; Li, J. PDB-wide collection of binding
data: Current status of the PDBbind database. Bioinformatics 2014, 31, 405–412.
77. Ballester, P.J.; Mitchell, J.B.O. A machine learning approach to predicting protein–ligand binding affinity
with applications to molecular docking. Bioinformatics 2010, 26, 1169–1175.
78. Jiménez, J.; Škalič, M.; Martínez-Rosell, G.; De Fabritiis, G. KDEEP: Protein-Ligand Absolute Binding
Affinity Prediction via 3D-Convolutional Neural Networks. J. Chem. Inf. Model. 2018, 58, 287–296.
79. Durrant, J.D.; McCammon, J.A. NNScore: A neural-network-based scoring function for the characterization
of protein-ligand complexes. J. Chem. Inf. Model. 2010, 50, 1865–1871.
80. Vreven, T.; Moal, I.H.; Vangone, A.; Pierce, B.G.; Kastritis, P.L.; Torchala, M.; Chaleil, R.; Jiménez-García,
B.; Bates, P.A.; Fernandez-Recio, J.; et al. Updates to the Integrated Protein–Protein Interaction Benchmarks:
Docking Benchmark Version 5 and Affinity Benchmark Version 2. J. Mol. Biol. 2015, 427, 3031–3041.
81. Koukos, P.I.; Faro, I.; van Noort, C.W.; Bonvin, A.M.J.J. A Membrane Protein Complex Docking Benchmark.
J. Mol. Biol. 2018, 430, 5246–5256.
82. Li, H.; Leung, K.S.; Wong, M.H.; Ballester, P.J. Improving autodock vina using random forest: The growing
accuracy of binding affinity prediction by the effective exploitation of larger data sets. Mol. Inform. 2015,
34, 115–126.
83. Ain, Q.U.; Aleksandrova, A.; Roessler, F.D.; Ballester, P.J. Machine-learning scoring functions to improve
structure-based binding affinity prediction and virtual screening. Wiley Interdiscip. Rev. Comput. Mol. Sci.
2015, 5, 405–424.
84. Irwin, J.J. Community benchmarks for virtual screening. J. Comput. Aided Mol. Des. 2008, 22, 193–199.
85. Kirchmair, J.; Markt, Æ.P.; Distinto, S.; Wolber, Æ.G. Evaluation of the performance of 3D virtual screening
protocols: RMSD comparisons, enrichment assessments, and decoy selection—What can we learn from
earlier mistakes? J. Comput. Aided Mol. Des. 2008, 22, 213–228.
86. Verdonk, M.L.; Berdini, V.; Hartshorn, M.J.; Mooij, W.T.M.; Murray, C.W.; Taylor, R.D.; Watson, P. Virtual
screening using protein-ligand docking: Avoiding artificial enrichment. J. Chem. Inf. Comput. Sci. 2004, 44,
793–806.
Int. J. Mol. Sci. 2019, 20, 4574 24 of 29

87. Good, A.C.; Oprea, T.I. Optimization of CAMD techniques 3. Virtual screening enrichment studies: A help
or hindrance in tool selection? J. Comput. Aided Mol. Des. 2008, 22, 169–178.
88. Lovell, T.; Chen, H.; Lyne, P.D.; Giordanetto, F.; Li, J. On Evaluating Molecular-Docking Methods for Pose
Prediction and Enrichment Factors. [J. Chem. Inf. Model. 46, 401−415 (2006)] by. J. Chem. Inf. Model. 2008,
48, 246–246.
89. Vogel, S.M.; Bauer, M.R.; Boeckler, F.M. DEKOIS: Demanding evaluation kits for objective in silico
screening—A versatile tool for benchmarking docking programs and scoring functions. J. Chem. Inf. Model.
2011, 51, 2650–2665.
90. Huang, N.; Shoichet, B.K.; Irwin, J.J. Benchmarking Sets for Molecular Docking Benchmarking Sets for
Molecular Docking. Society 2006, 49, 6789–6801.
91. Wallach, I.; Lilien, R. Virtual decoy sets for molecular docking benchmarks. J. Chem. Inf. Model. 2011, 51,
196–202.
92. Mysinger, M.M.; Carchia, M.; Irwin, J.J.; Shoichet, B.K. Directory of useful decoys, enhanced (DUD-E):
Better ligands and decoys for better benchmarking. J. Med. Chem. 2012, 55, 6582–6594.
93. Bauer, M.R.; Ibrahim, T.M.; Vogel, S.M.; Boeckler, F.M. Evaluation and optimization of virtual screening
workflows with DEKOIS 2.0—A public library of challenging docking benchmark sets. J. Chem. Inf. Model.
2013, 53, 1447–1462.
94. Rohrer, S.G.; Baumann, K. Maximum unbiased validation (MUV) data sets for virtual screening based on
PubChem bioactivity data. J. Chem. Inf. Model. 2009, 49, 169–184.
95. Gatica, E.A.; Cavasotto, C.N. Ligand and decoy sets for docking to G protein-coupled receptors. J. Chem.
Inf. Model. 2012, 52, 1–6.
96. Lagarde, N.; Ben Nasr, N.; Jérémie, A.; Guillemain, H.; Laville, V.; Labib, T.; Zagury, J.F.; Montes, M.
NRLiSt BDB, the manually curated nuclear receptors ligands and structures benchmarking database. J.
Med. Chem. 2014, 57, 3117–3125.
97. Xia, J.; Tilahun, E.L.; Kebede, E.H.; Reid, T.E.; Zhang, L.; Wang, X.S. Comparative modeling and
benchmarking data sets for human histone deacetylases and sirtuin families. J. Chem. Inf. Model. 2015, 55,
374–388.
98. Cereto-Massagué, A.; Guasch, L.; Valls, C.; Mulero, M.; Pujadas, G.; Garcia-Vallvé, S. DecoyFinder: An
easy-to-use python GUI application for building target-specific decoy sets. Bioinformatics 2012, 28, 1661–
1662.
99. Wang, L.; Pang, X.; Li, Y.; Zhang, Z.; Tan, W. RADER: A RApid DEcoy Retriever to facilitate decoy based
assessment of virtual screening. Bioinformatics 2017, 33, 1235–1237.
100. Fawcett, T. An introduction to ROC analysis. Pattern Recognit. Lett. 2006, 27, 861–874.
101. Triballeau, N.; Acher, F.; Brabet, I.; Pin, J.-P.; Bertrand, H.-O. Virtual Screening Workflow Development
Guided by the “Receiver Operating Characteristic” Curve Approach. Application to High-Throughput
Docking on Metabotropic Glutamate Receptor Subtype 4. J. Med. Chem. 2005, 48, 2534–2547.
102. Truchon, J.F.; Bayly, C.I. Evaluating virtual screening methods: Good and bad metrics for the “early
recognition” problem. J. Chem. Inf. Model. 2007, 47, 488–508.
103. Empereur-Mot, C.; Guillemain, H.; Latouche, A.; Zagury, J.F.; Viallon, V.; Montes, M. Predictiveness curves
in virtual screening. J. Cheminform. 2015, 7, doi:10.1186/s13321-015-0100-8.
104. Alghamedy, F.; Bopaiah, J.; Jones, D.; Zhang, X.; Weiss, H.L.; Ellingson, S.R. Incorporating Protein
Dynamics Through Ensemble Docking in Machine Learning Models to Predict Drug Binding. AMIA Jt.
Summits Transl. Sci. 2018, 2017, 26–34.
105. Sheridan, R.P.; Singh, S.B.; Fluder, E.M.; Kearsley, S.K. Protocols for Bridging the Peptide to Nonpeptide
Gap in Topological Similarity Searches. J. Chem. Inf. Comput. Sci. 2002, 41, 1395–1406.
106. 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.
1999, 42, 5100–5109.
107. Wang, R.; Wang, S. How does consensus scoring work for virtual library screening? An idealized computer
experiment. J. Chem. Inf. Comput. Sci. 2001, 41, 1422–1426.
108. Kang, L.; Li, H.; Jiang, H.; Wang, X.; Zheng, M.; Luo, J.; Zhang, H.; Liu, X. An effective docking strategy for
virtual screening based on multi-objective optimization algorithm. BMC Bioinform. 2009, 10, 58.
Int. J. Mol. Sci. 2019, 20, 4574 25 of 29

109. Nguyen, D.D.; Cang, Z.; Wu, K.; Wang, M.; Cao, Y.; Wei, G.W. Mathematical deep learning for pose and
binding affinity prediction and ranking in D3R Grand Challenges. J. Comput. Aided Mol. Des. 2018, 33, 71–
82.
110. Wang, R.; Lu, Y.; Wang, S. Comparative evaluation of 11 scoring functions for molecular docking. J. Med.
Chem. 2003, 46, 2287–2303.
111. Ren, X.; Shi, Y.-S.; Zhang, Y.; Liu, B.; Zhang, L.-H.; Peng, Y.-B.; Zeng, R. Novel Consensus Docking Strategy
to Improve Ligand Pose Prediction. J. Chem. Inf. Model. 2018, 58, 1662–1668.
112. Poli, G.; Martinelli, A.; Tuccinardi, T. Reliability analysis and optimization of the consensus docking
approach for the development of virtual screening studies. J. Enzyme Inhib. Med. Chem. 2016, 31, 167–173.
113. Tuccinardi, T.; Poli, G.; Romboli, V.; Giordano, A.; Martinelli, A. Extensive consensus docking evaluation
for ligand pose prediction and virtual screening studies. J. Chem. Inf. Model. 2014, 54, 2980–2986.
114. Houston, D.R.; Walkinshaw, M.D. Consensus docking: Improving the reliability of docking in a virtual
screening context. J. Chem. Inf. Model. 2013, 53, 384–390.
115. Plewczynski, D.; Łażniewski, M.; Grotthuss, M. Von; Rychlewski, L.; Ginalski, K. VoteDock: Consensus
docking method for prediction of protein-ligand interactions. J. Comput. Chem. 2011, 32, 568–581.
116. Perez-castillo, Y.; Sotomayor-burneo, S.; Jimenes-vargas, K.; Gonzalez-, M. CompScore: Boosting structure-
based virtual screening performance by incorporating docking scoring functions components into
consensus scoring. BioRxiv 2019, doi:10.1101/550590.
117. Onawole, A.T.; Kolapo, T.U.; Sulaiman, K.O.; Adegoke, R.O. Structure based virtual screening of the Ebola
virus trimeric glycoprotein using consensus scoring. Comput. Biol. Chem. 2018, 72, 170–180.
118. Aliebrahimi, S.; Karami, L.; Arab, S.S.; Montasser Kouhsari, S.; Ostad, S.N. Identification of Phytochemicals
Targeting c-Met Kinase Domain using Consensus Docking and Molecular Dynamics Simulation Studies.
Cell Biochem. Biophys. 2017, 76, 135–145.
119. Li, D.D.; Meng, X.F.; Wang, Q.; Yu, P.; Zhao, L.G.; Zhang, Z.P.; Wang, Z.Z.; Xiao, W. Consensus scoring
model for the molecular docking study of mTOR kinase inhibitor. J. Mol. Graph. Model. 2018, 79, 81–87.
120. Oda, A.; Tsuchida, K.; Takakura, T.; Yamaotsu, N.; Hirono, S. Comparison of consensus scoring strategies
for evaluating computational models of protein-ligand complexes. J. Chem. Inf. Model. 2006, 46, 380–391.
121. Chaput, L.; Martinez-Sanz, J.; Quiniou, E.; Rigolet, P.; Saettel, N.; Mouawad, L. VSDC: A method to
improve early recognition in virtual screening when limited experimental resources are available. J.
Cheminform. 2016, 8, doi:10.1186/s13321-016-0112-z.
122. Mavrogeni, M.E.; Pronios, F.; Zareifi, D.; Vasilakaki, S.; Lozach, O.; Alexopoulos, L.; Meijer, L.;
Myrianthopoulos, V.; Mikros, E. A facile consensus ranking approach enhances virtual screening
robustness and identifies a cell-active DYRK1α inhibitor. Future Med. Chem. 2018, 10, 2411–2430.
123. Zhan, W.; Li, D.; Che, J.; Zhang, L.; Yang, B.; Hu, Y.; Liu, T.; Dong, X. Integrating docking scores, interaction
profiles and molecular descriptors to improve the accuracy of molecular docking: Toward the discovery of
novel Akt1 inhibitors. Eur. J. Med. Chem. 2014, 75, 11–20.
124. Ericksen, S.S.; Wu, H.; Zhang, H.; Michael, L.A.; Newton, M.A.; Hoffmann, F.M.; Wildman, S.A. Machine
Learning Consensus Scoring Improves Performance Across Targets in Structure-Based Virtual Screening.
J. Chem. Inf. Model. 2017, 57, 1579–1590.
125. Teramoto, R.; Fukunishi, H. Supervised consensus scoring for docking and virtual screening. J. Chem. Inf.
Model. 2007, 47, 526–534.
126. Wang, W.; He, W.; Zhou, X.; Chen, X. Optimization of molecular docking scores with support vector rank
regression. Proteins Struct. Funct. Bioinform. 2013, 81, 1386–1398.
127. Yang, J.M.; Hsu, D.F. Consensus scoring criteria in structure-based virtual screening. Emerg. Inf. Technol.
Conf. 2005 2005, 2005, 165–167.
128. Liu, S.; Fu, R.; Zhou, L.-H.; Chen, S.-P. Application of Consensus Scoring and Principal Component
Analysis for Virtual Screening against β-Secretase (BACE-1). PLoS ONE 2012, 7, e38086.
129. Mokrani, E.H.; Bensegueni, A.; Chaput, L.; Beauvineau, C.; Djeghim, H.; Mouawad, L. Identification of
New Potent Acetylcholinesterase Inhibitors Using Virtual Screening and In Vitro Approaches. Mol. Inform.
2019, 38, 1800118.
130. Russo Spena, C.; De Stefano, L.; Poli, G.; Granchi, C.; El Boustani, M.; Ecca, F.; Grassi, G.; Grassi, M.;
Canzonieri, V.; Giordano, A.; et al. Virtual screening identifies a PIN1 inhibitor with possible antiovarian
cancer effects. J. Cell. Physiol. 2019, doi:10.1002/jcp.28224.
Int. J. Mol. Sci. 2019, 20, 4574 26 of 29

131. Mouawad, N.; Jha, V.; Poli, G.; Granchi, C.; Rizzolio, F.; Caligiuri, I.; Minutolo, F.; Lapillo, M.; Tuccinardi,
T.; Macchia, M. Computationally driven discovery of phenyl(piperazin-1-yl) methanone derivatives as
reversible monoacylglycerol lipase (MAGL) inhibitors. J. Enzyme Inhib. Med. Chem. 2019, 34, 589–596.
132. Damm-Ganamet, K.L.; Dunbar, J.B.; Ahmed, A.; Esposito, E.X.; Stuckey, J.A.; Gestwicki, J.E.;
Chinnaswamy, K.; Delproposto, J.; Smith, R.D.; Carlson, H.A.; et al. CSAR Data Set Release 2012: Ligands,
Affinities, Complexes, and Docking Decoys. J. Chem. Inf. Model. 2013, 53, 1842–1852.
133. Walters, W.P.; Liu, S.; Chiu, M.; Shao, C.; Rudolph, M.G.; Burley, S.K.; Gilson, M.K.; Feher, V.A.; Gaieb, Z.;
Kuhn, B.; et al. D3R Grand Challenge 2: Blind prediction of protein–ligand poses, affinity rankings, and
relative binding free energies. J. Comput. Aided Mol. Des. 2017, 32, 1–20.
134. Nevins, N.; Yang, H.; Walters, W.P.; Ameriks, M.K.; Parks, C.D.; Gilson, M.K.; Gaieb, Z.; Lambert, M.H.;
Shao, C.; Chiu, M.; et al. D3R Grand Challenge 3: Blind prediction of protein–ligand poses and affinity
rankings. J. Comput. Aided Mol. Des. 2019, 33, 1–18.
135. Shuker, S.B.; Hajduk, P.J.; Meadows, R.P.; Fesik, S.W. Discovering High-Affinity Ligands for Proteins: SAR
by NMR. Science 1996, 274, 1531–1534.
136. Romasanta, A.K.S.; van der Sijde, P.; Hellsten, I.; Hubbard, R.E.; Keseru, G.M.; van Muijlwijk-Koezen, J.;
de Esch, I.J.P. When fragments link: A bibliometric perspective on the development of fragment-based drug
discovery. Drug Discov. Today 2018, 23, 1596–1609.
137. Erlanson, D.A. Introduction to fragment-based drug discovery. Top. Curr. Chem. 2012, 317, 1–32.
138. Hann, M.M.; Leach, A.R.; Harper, G. Molecular Complexity and Its Impact on the Probability of Finding
Leads for Drug Discovery. J. Chem. Inf. Comput. Sci. 2001, 41, 856–864.
139. Leach, A.R.; Hann, M.M. Molecular complexity and fragment-based drug discovery: Ten years on. Curr.
Opin. Chem. Biol. 2011, 15, 489–496.
140. Fink, T.; Raymond, J.L. Virtual exploration of the chemical universe up to 11 atoms of C, N, O, F: Assembly
of 26.4 million structures (110.9 million stereoisomers) and analysis for new ring systems, stereochemistry,
physicochemical properties, compound classes, and drug discovery. J. Chem. Inf. Model. 2007, 47, 342–353.
141. Lyu, J.; Irwin, J.J.; Roth, B.L.; Shoichet, B.K.; Levit, A.; Wang, S.; Tolmachova, K.; Singh, I.; Tolmachev, A.A.;
Che, T.; et al. Ultra-large library docking for discovering new chemotypes. Nature 2019, 566, 224.
142. Scott, D.E.; Coyne, A.G.; Hudson, S.A.; Abell, C. Fragment-based approaches in drug discovery and
chemical biology. Biochemistry 2012, 51, 4990–5003.
143. Blundell, T.L.; Jhoti, H.; Abell, C. High-throughput crystallography for lead discovery in drug design. Nat.
Rev. Drug Discov. 2002, 1, 45–54.
144. Hopkins, A.L.; Groom, C.R.; Alex, A. Ligand efficiency: A useful metric for lead selection. Drug Discov.
Today 2004, 9, 430–431.
145. Abad-Zapatero, C.; Metz, J.T. Ligand efficiency indices as guideposts for drug discovery. Drug Discov.
Today 2005, 10, 464–469.
146. Reynolds, C.H.; Bembenek, S.D.; Tounge, B.A. The role of molecular size in ligand efficiency. Bioorg. Med.
Chem. Lett. 2007, 17, 4258–4261.
147. Schultes, S.; De Graaf, C.; Haaksma, E.E.J.; De Esch, I.J.P.; Leurs, R.; Krämer, O. Ligand efficiency as a guide
in fragment hit selection and optimization. Drug Discov. Today Technol. 2010, 7, 157–162.
148. Congreve, M.; Carr, R.; Murray, C.; Jhoti, H. A ‘Rule of Three’ for fragment-based lead discovery? Recent.
Drug Discov. Today 2003, 8, 876–877.
149. Jhoti, H.; Williams, G.; Rees, D.C.; Murray, C.W. The “rule of three” for fragment-based drug discovery:
Where are we now? Nat. Rev. Drug Discov. 2013, 12, 644–644.
150. Lipinski, C.A.; Lombardo, F.; Dominy, B.W.; Feeney, P.J. Experimental and computational approaches to
estimate solubility and permeability in drug discovery and development settings. Adv. Drug Deliv. Rev.
2001, 46, 3–26.
151. Morley, A.D.; Pugliese, A.; Birchall, K.; Bower, J.; Brennan, P.; Brown, N.; Chapman, T.; Drysdale, M.;
Gilbert, I.H.; Hoelder, S.; et al. Fragment-based hit identification: Thinking in 3D. Drug Discov. Today 2013,
18, 1221–1227.
152. Verheij, H.J. Leadlikeness and structural diversity of synthetic screening libraries. Mol. Divers. 2006, 10,
377–388.
153. Fischer, M.; Hubbard, R.E. Fragment-based ligand discovery. Mol. Interv. 2009, 9, 22–30.
154. Schuffenhauer, A.; Ruedisser, S.; Jahnke, W.; Marzinzik, A.; Selzer, P.; Jacoby, E. Library Design for
Fragment Based Screening. Curr. Top. Med. Chem. 2005, 5, 751–762.
Int. J. Mol. Sci. 2019, 20, 4574 27 of 29

155. Lewell, X.Q.; Judd, D.B.; Watson, S.P.; Hann, M.M. RECAP—Retrosynthetic Combinatorial Analysis
Procedure: A powerful new technique for identifying privileged molecular fragments with useful
applications in combinatorial chemistry. J. Chem. Inf. Comput. Sci. 1998, 38, 511–522.
156. Prescher, H.; Koch, G.; Schuhmann, T.; Ertl, P.; Bussenault, A.; Glick, M.; Dix, I.; Petersen, F.; Lizos, D.E.
Construction of a 3D-shaped, natural product like fragment library by fragmentation and diversification
of natural products. Bioorg. Med. Chem. 2017, 25, 921–925.
157. Chen, Y.; Shoichet, B.K. Molecular docking and ligand specificity in fragment-based inhibitor discovery.
Nat. Chem. Biol. 2009, 5, 358–364.
158. Fjellström, O.; Akkaya, S.; Beisel, H.G.; Eriksson, P.O.; Erixon, K.; Gustafsson, D.; Jurva, U.; Kang, D.; Karis,
D.; Knecht, W.; et al. Creating novel activated factor XI inhibitors through fragment based lead generation
and structure aided drug design. PLoS ONE 2015, 10, e0113705.
159. Park, H.; Shin, Y.; Kim, J.; Hong, S. Application of Fragment-Based de Novo Design to the Discovery of
Selective Picomolar Inhibitors of Glycogen Synthase Kinase-3 Beta. J. Med. Chem. 2016, 59, 9018–9034.
160. Wang, R.; Gao, Y.; Lai, L. LigBuilder: A Multi-Purpose Program for Structure-Based Drug Design. J. Mol.
Model. 2004, 6, 498–516.
161. Zhao, H.; Gartenmann, L.; Dong, J.; Spiliotopoulos, D.; Caflisch, A. Discovery of BRD4 bromodomain
inhibitors by fragment-based high-throughput docking. Bioorg. Med. Chem. Lett. 2014, 24, 2493–2496.
162. Rudling, A.; Gustafsson, R.; Almlöf, I.; Homan, E.; Scobie, M.; Warpman Berglund, U.; Helleday, T.;
Stenmark, P.; Carlsson, J. Fragment-Based Discovery and Optimization of Enzyme Inhibitors by Docking
of Commercial Chemical Space. J. Med. Chem. 2017, 60, 8160–8169.
163. Hernandez, J.; Hoffer, L.; Coutard, B.; Querat, G.; Roche, P.; Morelli, X.; Decroly, E.; Barral, K. Optimization
of a fragment linking hit toward Dengue and Zika virus NS5 methyltransferases inhibitors. Eur. J. Med.
Chem. 2019, 161, 323–333.
164. Akabayov, S.R.; Richardson, C.C.; Arthanari, H.; Akabayov, B.; Ilic, S.; Wagner, G. Identification of DNA
primase inhibitors via a combined fragment-based and virtual screening. Sci. Rep. 2016, 6, 36322.
165. Amaning, K.; Lowinski, M.; Vallee, F.; Steier, V.; Marcireau, C.; Ugolini, A.; Delorme, C.; Foucalt, F.;
McCort, G.; Derimay, N.; et al. The use of virtual screening and differential scanning fluorimetry for the
rapid identification of fragments active against MEK1. Bioorg. Med. Chem. Lett. 2013, 23, 3620–3626.
166. Barelier, S.; Eidam, O.; Fish, I.; Hollander, J.; Figaroa, F.; Nachane, R.; Irwin, J.J.; Shoichet, B.K.; Siegal, G.
Increasing chemical space coverage by combining empirical and computational fragment screens. ACS
Chem. Biol. 2014, 9, 1528–1535.
167. Adams, M.; Kobayashi, T.; Lawson, J.D.; Saitoh, M.; Shimokawa, K.; Bigi, S.V.; Hixon, M.S.; Smith, C.R.;
Tatamiya, T.; Goto, M.; et al. Fragment-based drug discovery of potent and selective MKK3/6 inhibitors.
Bioorg. Med. Chem. Lett. 2016, 26, 1086–1089.
168. Darras, F.H.; Pockes, S.; Huang, G.; Wehle, S.; Strasser, A.; Wittmann, H.J.; Nimczick, M.; Sotriffer, C.A.;
Decker, M. Synthesis, biological evaluation, and computational studies of Tri- and tetracyclic nitrogen-
bridgehead compounds as potent dual-acting AChE inhibitors and h H3 receptor antagonists. ACS Chem.
Neurosci. 2014, 5, 225–242.
169. He, Y.; Guo, X.; Yu, Z.H.; Wu, L.; Gunawan, A.M.; Zhang, Y.; Dixon, J.E.; Zhang, Z.Y. A potent and selective
inhibitor for the UBLCP1 proteasome phosphatase. Bioorg. Med. Chem. 2015, 23, 2798–2809.
170. Bishop, C.M. Pattern Recognition and Machine Learning; Springer: New York, NY, USA, 2007; ISBN 978-0-
387-31073-2.
171. Ashtawy, H.M.; Mahapatra, N.R. A comparative assessment of predictive accuracies of conventional and
machine learning scoring functions for protein-ligand binding affinity prediction. IEEE/ACM Trans.
Comput. Biol. Bioinform. 2015, 12, 335–347.
172. Ashtawy, H.M.; Mahapatra, N.R. Machine-learning scoring functions for identifying native poses of
ligands docked to known and novel proteins. BMC Bioinform. 2015, 16, doi:10.1186/1471-2105-16-S6-S3.
173. Hassan, M.; Mogollón, D.C.; Fuentes, O. DLSCORE: A Deep Learning Model for Predicting Protein-Ligand
Binding Affinities. ChemRxiv 2018, 13, 53.
174. Ouyang, X.; Handoko, S.D.; Kwoh, C.K. Cscore: A Simple Yet Effective Scoring Function for Protein–
Ligand Binding Affinity Prediction Using Modified Cmac Learning Architecture. J. Bioinform. Comput. Biol.
2011, 9, 1–14.
Int. J. Mol. Sci. 2019, 20, 4574 28 of 29

175. Kinnings, S.L.; Liu, N.; Tonge, P.J.; Jackson, R.M.; Xie, L.; Bourne, P.E. A machine learning-based method
to improve docking scoring functions and its application to drug repurposing. J. Chem. Inf. Model. 2011, 51,
408–419.
176. Hsin, K.Y.; Ghosh, S.; Kitano, H. Combining machine learning systems and multiple docking simulation
packages to improve docking prediction reliability for network pharmacology. PLoS ONE 2013, 8, e83922.
177. Pereira, J.C.; Caffarena, E.R.; Dos Santos, C.N. Boosting Docking-Based Virtual Screening with Deep
Learning. J. Chem. Inf. Model. 2016, 56, 2495–2506.
178. Pason, L.P.; Sotriffer, C.A. Empirical Scoring Functions for Affinity Prediction of Protein-ligand Complexes.
Mol. Inform. 2016, 35, 541–548.
179. Silva, C.G.; Simoes, C.J.V.; Carreiras, P.; Brito, R.M.M. Enhancing Scoring Performance of Docking-Based
Virtual Screening Through Machine Learning. Curr. Bioinform. 2016, 11, 408–420.
180. Korkmaz, S.; Zararsiz, G.; Goksuluk, D. MLViS: A web tool for machine learning-based virtual screening
in early-phase of drug discovery and development. PLoS ONE 2015, 10, e0124600.
181. Springer, C.; Adalsteinsson, H.; Young, M.M.; Kegelmeyer, P.W.; Roe, D.C. PostDOCK: A Structural,
Empirical Approach to Scoring Protein Ligand Complexes. J. Med. Chem. 2005, 48, 6821–6831.
182. Ashtawy, H.M.; Mahapatra, N.R. Task-Specific Scoring Functions for Predicting Ligand Binding Poses and
Affinity and for Screening Enrichment. J. Chem. Inf. Model. 2018, 58, 119–133.
183. Imrie, F.; Bradley, A.R.; Van Der Schaar, M.; Deane, C.M. Protein Family-Specific Models Using Deep
Neural Networks and Transfer Learning Improve Virtual Screening and Highlight the Need for More Data.
J. Chem. Inf. Model. 2018, 58, 2319–2330.
184. Wang, Y.; Guo, Y.; Kuang, Q.; Pu, X.; Ji, Y.; Zhang, Z.; Li, M. A comparative study of family-specific protein-
ligand complex affinity prediction based on random forest approach. J. Comput. Aided Mol. Des. 2015, 29,
349–360.
185. Wójcikowski, M.; Ballester, P.J.; Siedlecki, P. Performance of machine-learning scoring functions in
structure-based virtual screening. Sci. Rep. 2017, 7, 46710.
186. Cao, Y.; Li, L. Improved protein-ligand binding affinity prediction by using a curvature-dependent surface-
area model. Bioinformatics 2014, 30, 1674–1680.
187. Yuriev, E.; Ramsland, P.A. Latest developments in molecular docking: 2010–2011 in review. J. Mol. Recognit.
2013, 26, 215–239.
188. Guedes, I.A.; Pereira, F.S.S.; Dardenne, L.E. Empirical scoring functions for structure-based virtual
screening: Applications, critical aspects, and challenges. Front. Pharmacol. 2018, 9, 1089.
189. Li, L.; Wang, B.; Meroueh, S.O. Support Vector Regression Scoring of Receptor–Ligand Complexes for
Rank-Ordering and Virtual Screening of Chemical Libraries. J. Chem. Inf. Model. 2011, 51, 2132–2138.
190. Guyon, I.; Elisseeff, A. An Introduction to Variable and Feature Selection. J. Mach. Learn. Res. 2011, 3, 1157–
1182.
191. Koppisetty, C.A.K.; Frank, M.; Kemp, G.J.L.; Nyholm, P.G. Computation of binding energies including
their enthalpy and entropy components for protein-ligand complexes using support vector machines. J.
Chem. Inf. Model. 2013, 53, 2559–2570.
192. Liu, Q.; Kwoh, C.K.; Li, J. Binding affinity prediction for protein-ligand complexes based on β contacts and
B factor. J. Chem. Inf. Model. 2013, 53, 3076–3085.
193. Ballester, P.J.; Schreyer, A.; Blundell, T.L. Does a More Precise Chemical Description of Protein—Ligand
Complexes Lead to More Accurate Prediction of Binding Affinity? J. Chem. Inf. Model. 2014, 54, 944–955.
194. Kundu, I.; Paul, G.; Banerjee, R. A machine learning approach towards the prediction of protein–ligand
binding affinity based on fundamental molecular properties. RSC Adv. 2018, 8, 12127–12137.
195. Srinivas, R.; Klimovich, P.V.; Larson, E.C. Implicit-descriptor ligand-based virtual screening by means of
collaborative filtering. J. Cheminform. 2018, 10, 56.
196. Ragoza, M.; Hochuli, J.; Idrobo, E.; Sunseri, J.; Koes, D.R. Protein-Ligand Scoring with Convolutional
Neural Networks. J. Chem. Inf. Model. 2017, 57, 942–957.
197. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. Convolutional Neural Networks ImageNet Classification with
Deep Convolutional Neural Network. Commun. ACM 2017, 60, doi:10.1145/3065386.
198. Khamis, M.A.; Gomaa, W.; Ahmed, W.F. Machine learning in computational docking. Artif. Intell. Med.
2015, 63, 135–152.
199. Sieg, J.; Flachsenberg, F.; Rarey, M. In the need of bias control: Evaluation of chemical data for Machine
Learning Methods in Virtual Screening. J. Chem. Inf. Model. 2019, 59, 947–961.
Int. J. Mol. Sci. 2019, 20, 4574 29 of 29

200. Durrant, J.D.; Carlson, K.E.; Martin, T.A.; Offutt, T.L.; Mayne, C.G.; Katzenellenbogen, J.A.; Amaro, R.E.
Neural-Network Scoring Functions Identify Structurally Novel Estrogen-Receptor Ligands. J. Chem. Inf.
Model. 2015, 55, 1953–1961.
201. Pires, D.E.V.; Ascher, D.B. CSM-lig: A web server for assessing and comparing protein-small molecule
affinities. Nucleic Acids Res. 2016, 44, W557–W561.
202. Zilian, D.; Sotriffer, C.A. SFCscore RF: A Random Forest-Based Scoring Function for Improved Affinity
Prediction of Protein–Ligand Complexes. J. Chem. Inf. Model. 2013, 53, 1923–1933.
203. Li, G.-B.; Yang, L.-L.; Wang, W.-J.; Li, L.-L.; Yang, S.-Y. ID-Score: A New Empirical Scoring Function Based
on a Comprehensive Set of Descriptors Related to Protein–Ligand Interactions. J. Chem. Inf. Model. 2013, 53,
592–600.

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons
Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).