Combinatorial Chemistry & High Throughput Screening, 2011, 14, 889-897 889

Integration of Virtual and High Throughput Screening in Lead Discovery

Settings
Tímea Polgár1 and György M. Keser*,1,2

1
Budapest University of Technology and Economics, Szt. Gellért tér 4, 1111 Budapest, Hungary
2
Gedeon Richter Plc, P.O. Box 27, H-1475 Budapest, Hungary

Abstract: In the last decade mass screening strategies became the main source of leads in drug discovery settings.
Although high throughput (HTS) and virtual screening (VS) realize the same concept the different nature of these lead
discovery strategies (experimental vs theoretical) results that they are typically applied separately. The majority of drug
leads are still identified by hit-to-lead optimization of screening hits. Structural information on the target as well as on
bound ligands, however, make structure-based and ligand-based virtual screening available for the identification of
alternative chemical starting points. Although, the two techniques have rarely been used together on the same target, here
we review the existing prominent studies on their true integration. Various approaches have been shown to apply the
combination of HTS and VS and to better use them in lead generation. Although several attempts on their integration have
only been considered at a conceptual level, there are numerous applications underlining its relevance that early-stage
pharmaceutical drug research could benefit from a combined approach.
Keywords: Virtual screening, high throughput screening, integration, parallel screening, focused screening, sequential
screening.

INTRODUCTION optimization of the protocols used for large scale prediction

Identification of suitable chemical starting points and
viable leads is one of the key tasks of early phase drug
discovery. Historically, lead discovery was mainly supported
by in vivo experiments usually resulting in compounds with
acceptable efficacy and suitable pharmacokinetic profile in
animal models. However, market expectations of pharma
investors facilitated the application of other approaches such
as in vitro screening or computer-aided drug discovery
(CADD) methods [1].
According to a recent estimation experimental screening
methods together provide 65% of drug leads and account for
approximately 14% of the preclinical budget [2]. In vitro
assays serve the basis of structure-activity relationships that
drive medicinal chemistry during active-to-hit and hit-to-lead
processes. Dramatic developments in molecular biology,
detection methods, computer technology, and automation
and miniaturization made HTS a characteristic tool of lead
discovery. For today, it is clear that library design and well-
designed in vitro assays lead to higher hit rates reflecting a
current trend to ‘re-rationalize’ drug discovery research.
Random screening should therefore be replaced by
methodologies that combine the capacity of HTS approaches
and the rational basis of CADD techniques [3].
Application of the HTS technology requires the selection
of the target, the development of the assay and the
availability of the screening library in a suitable format. In
this respect virtual screening techniques can be considered as
in silico analogues of in vitro HTS technologies. After
defining the target, VS also requires the development and
*Address correspondence to this author at the Gedeon Richter Plc, P.O. Box
27, H-1475 Budapest, Hungary; Tel: +36-1-4314605; Fax: +36-1-4326002;
E-mail: gy.keseru@richter.hu
of biological affinity. In addition, HTS and VS have another
common conceptual framework; both methods have limited
accuracy that is compensated by the number of compounds
investigated. VS and HTS can therefore be considered as
classification techniques that separate actives from inactives
rather than identifying validated hits. Actives picked up by
HTS or VS should therefore be experimentally validated
before starting hit-to-lead process. Therefore, HTS and VS
are not considered as competitive disciplines. However in
reality, HTS and VS methods are still less integrated than it
is expected. In addition to organizational and disciplinary
aspects one of the reasons can be cultural based on the
different viewpoints of experimentalists and theoreticians.
During the last decade, various in silico methods were
applied for the analysis of large screening datasets and for
the knowledge extraction. Moreover predictive computat-
ional models were attempted to build on the extracted
knowledge aiming at using the two complementary methods
in practice. In addition to integrated VS and HTS solutions,
these methods, of which intended use is to improve hit rates
and accelerate pharmaceutical research are discussed in this
review.
VIRTUAL SCREENING
Virtual screening (VS) has emerged as an inexpensive
and straightforward method for identifying primary actives
or even lead molecules. It can be either realized as structure-
based screening approach (also called docking) or a ligand-
based one considering chemical structure similarity of small
molecules during database screening. Here we restrict
ourselves to discuss only the most relevant methods in
virtual screening (Fig. 1) and point to recent reviews as
references.

1386-2073/11 $58.00+.00 © 2011 Bentham Science Publishers

890 Combinatorial Chemistry & High Throughput Screening, 2011, Vol. 14, No. 10
N O
HN
N
N
Cl
Cl
2D
N
N
2D fingerprint substructure
Fig. (1). Schematic representation of Virtual Screening Approaches.
Database filtering. One of the best known early
physicochemical filters, the rule of five (ROF) was
introduced by Lipinski et al. [4]. It is based on the
distributions of easily accessible and interpretable
physicochemical properties of known oral drugs. These
properties define the so-called drug-like chemistry space
where compounds face with no serious ADME (absorption,
distribution, metabolism, and elimination) problems. Later,
Hann [5] and Oprea [6] introduced the concept of lead-
likeness after analysing lead molecules and corresponding
drugs. Comparative studies on leads and drugs revealed
lower complexity of lead molecules. Nevertheless, it should
be emphasized that limits applied for both drug-like and
lead-like compounds are based on statistical analyses on
known examples. Although these simple filters demonstrated
significant classification accuracy when discriminating drugs
and non-drugs, leads and drugs their nature is mainly
empirical. Unstable, reactive, toxic or otherwise unsuitable
compounds can simply be removed using functional group
filters [7, 8].
Ligand-based screening [9]. Similarity searching is one
of the simplest methodologies of ligand-based virtual
screening when screening databases against a query using
chemical similarity principles. During a similarity search the
query molecule is compared to members of a database and a
measure is calculated quantifying the similarity between the
query and each molecule of the database [10]. For 2D
similarity searches the query structure and members of the
screened database are typically represented by molecular
fingerprints that encode molecular structure and properties in
binary format (Fig. 1). A given number of bits in the bit
string detect the presence or absence of molecular fragments
Polgár and Keser
docking
CH
3
Cl
3D
Trp255
Tyr275 Asp366-Lys192
Val196
A
Phe278 Phe170
Trp279
C D E
Leu387
B
Met384
Trp356
Phe200
3D pharmacophore
shape similarity
phramacophore fingerprint
or connectivity pattern and pharmacophore, can encode
descriptor value ranges or binary transformed descriptors.
Analysis of molecular similarity is based on the quantitative
determination of the overlap between fingerprints of the
query structure and all database members. Since descriptors
of a given molecule can be considered as a vector of real or
binary attributes most of the similarity measures are derived
as vectorial distances. If we consider small molecules as
spatial objects virtual screening can also be performed using
3D queries. 3D similarity searches utilize 3D information
including shape, steric and electrostatic properties obtained
for the query molecule and the database screened. High
throughput molecular alignment techniques of this kind
superpose all database entries onto the query molecule.
FlexS, one of the most popular approaches, keeps the query
rigid and considers test molecules as flexible when offering
several alternative superpositions each scored and ranked by
similarity [11, 12]. ROCS and ChemAxon’s 3D performs a
shape based superposition of the molecular volume
combining 2D and 3D similarity scoring [13, 14].
Pharmacophore is an arrangement of steric and electrostatic
features in the three dimensional space that are crucial for
the biological action. Structure-based pharmacophores are
typically explored by analyzing binding site interactions
formed between ligands and protein. Without having
structural information ligand-based pharmacophore can be
developed using a set of active compounds. Pharmacophore
searches use at least one conformation of database
compounds that is expected to be similar to the bioactive
conformation.
Structure-based virtual screening [15, 16]. Functions of
drug molecules and protein targets are regulated by the
Integration of Virtual and High Throughput Screening Combinatorial Chemistry & High Throughput Screening, 2011, Vol. 14, No. 10 891
principles of molecular recognition. The rational drug design
requires the understanding of molecular recognition in terms
of structure and energetics [17] that could be
straightforwardly realized by structure-based virtual
screening tools. Structure-based virtual screening starts with
obtaining the coordinates of a protein. X-ray crystallographic
or NMR 3D protein structures are used, but homology
models can also be applied [18]. Within the protein
structures active site has to be defined because scanning the
entire surface of the protein would hardly be feasible with
most of the currently used docking algorithms. Another
factor should also be considered, the conformational
flexibility of the target, as commonly used docking methods
are only able to consider flexibility to a limited extent.
Therefore the selection of the target structure is crucial;
binding sites should be as representative as possible,
mimicking a real and general binding conformation [19].
Moreover, ligand flexibility is also important and considered
using one of the two generally applied strategies: whole
molecule and fragment-based approaches. Historically, the
DOCK algorithm addressed rigid body docking using a
geometric matching algorithm to superimpose ligand
conformations onto a negative image of the binding pocket.
While fragment based methods provide an alternative
solution for the docking problem; molecules are dissected
into fragments that can be docked individually, either
separately or incrementally [20].
The interaction between protein and ligand is typically a
reversible equilibrium reaction usually characterised by the
free energy of binding. Binding free energy values with
reasonable accuracy can only be calculated requiring large
computational resources. Evaluation of thousands of ligands
requires a computationally cheap estimate of the binding free
energy. This goal could only be achieved by scoring schemes
that involves force filed-based, empirical and knowledge-
based methods [20]. All fast scoring functions share multiple
deficiencies. First of all most of them are fitted to or derived
from some kind of experimental data. Consequently, the
functions reflect the accuracy of these measurements.
Molecular size can influence the scores as well, generally the
larger the molecule is the better the score is, however
biological measurements do not support this observation.
Since scoring functions are derived from X-ray structures
only the favourable interactions are rewarded but
unfavourable interactions are not penalized because
information from the crystal structures cannot be obtained.
Uncertainties in the protonation states and the involvement
of water in ligand binding complicate scoring further. The
performance of scoring functions can be improved using
consensus scoring schemes in some cases. The disadvantages
and advantages of the various scores might provide a
combination, which is able to describe the main
characteristics of the protein-ligand binding [21].
HIGH THROUGHPUT SCREENING
High throughput screening is a key link in the chain
comprising the industrialized drug discovery paradigm.
Today, many pharmaceutical companies are screening
generally hundreds of thousands or more compounds per
screen to produce typically several hundreds to thousands of
hits. On average, some of these might be developed to lead
compound series. In terms of definition, high throughput
screening can be considered the process in which batches of
compounds are tested for binding affinity or functional
activity against the target. Test compounds act as inhibitors
of target enzymes, as competitors for binding of a natural
ligand to its receptor, as agonists, antagonists or allosteric
modulators for receptor-mediated intracellular processes, and
so forth. High throughput screening seeks to screen large
numbers of compounds rapidly and in a massively parallel
setup [22].
Primary positive high throughput screening results are
usually called actives. Activity, purity and identity of the
primary actives should first be confirmed and the resulting
hits are then subjected to detailed characterization. After this
second level of triage, hits are optimized in terms of in vitro
potency and ADME profile to become lead compounds [23].
The pharmaceutical industry currently has a pressing
need for improvement in high throughput screening strategy.
Although the industry has a seemingly insatiable appetite for
new lead compounds, it is also under continuing pressure to
reduce the costs of discovery and development.
High throughput screening instruments, assays, and services
have emerged as a significant growth market. From the field’s
origins with home-brew tests and generic research
instrumentation, high throughput screening has become an
increasingly sophisticated and important element in the
armamentarium of the drug discoverer. At the beginning of this
century the discovery process seemed to be on a steep growth
curve both with respect to the number of targets and compounds
to be screened and the complexity of the assays required.
Increased numbers of targets and compounds called for greater
parallelism and/or increased throughput in screens.
Furthermore, the expense and scarcity of targets and compounds
have driven the screening technology toward smaller assay
volumes through miniaturization [24]. At the end of the first
decade, however, an opposite trend became evident. Analyzing
lead generation strategies one can conclude that traditional
literature-based medicinal chemistry approaches and fragment
screening replace resource intensive random HTS campaigns. In
addition to the decreasing role of large scale screening in lead
generation recent HTS strategies are focused to screen smaller
compound collections with optimized content of chemical
information. The quality of the screening deck is of primary
importance generating high quality data. Furthermore, focused
library approaches are used more and more extensively to
improve the hit rate of the screening campaigns. It is important
to note, that most of the limitations of large scale random
screening could be overcome by integrating virtual screening
approaches. The demonstrated success of virtual screening
technologies in maximizing the chemical information content of
compound libraries, in reducing false positive and false negative
rates of experimental screening and finally in the design of
target family focused libraries are all supporting the integration
of virtual and experimental screening.
INTEGRATION OF VIRTUAL AND HIGH
THROUGHPUT SCREENING
Integration of virtual chemistry and screening realises
how a continuous information exchange between different
areas could lead to chemical libraries of probes with more
892 Combinatorial Chemistry & High Throughput Screening, 2011, Vol. 14, No. 10
desirable chemical and biological properties (Fig. 2). The
current trend in the pharmaceutical industry is to integrate
computational and experimental technologies early in the
drug discovery process. For instance, it is considered that a
better and earlier utilization of information, that is either
chemical or biological, would lead to chemical libraries with
more desirable chemical and biological properties. This can
be done by integrating information from different areas that
include: (i) analysis of the gene or protein family for target
selection; (ii) gene-structure-function studies by structural
biology approaches on the target; (iii) generation of virtual
and physical libraries by medicinal chemistry; (iv)
absorption, distribution, metabolism, excretion and toxicity
evaluation (ADME/ Tox) of virtual and physical libraries;
(v) VS of virtual and physical libraries and experimental
screening of virtual actives to identify chemical probes.
target selection
structural biology
ADMETox
small molecules
medicinal chemistry
virtual screening
chemical probe
Fig. (2). General strategy for the integration. Abbreviation:
ADMETox: absorption, distribution, metabolism, excretion and
toxicity.
The primary objective of HTS and VS is to derive actives
that could be realized by experimental and computational
approaches, respectively. The computational strategy applied
in VS suggests that it could be more effective in time,
resource, cost and hit-rates considering only experimentally
tested compounds. Moreover it does not require the physical
availability of the screening library and therefore it could
explore a significantly larger part of the drug-like chemistry
space. In spite of these advantages comparative studies
demonstrated that these approaches are rather
complementary than competitive. In one hand, VS
techniques could be useful when designing target specific
screening libraries. On the other hand, distinct structural
classes identified by VS and HTS are clearly the suggest
complementarities between VS and HTS.
COMPLEMENTARITIES IN DATA ANALYSIS
Complementary nature of HTS and VS are well used
during HTS data analysis where experimental knowledge is
extracted and used for predictive model building. These
models then can further be applied for data mining. Many of
the promising actives can never be validated, these are false
positives. Actives identified by HTS or VS have therefore to
Polgár and Keser
be experimentally validated before starting the hit-to-lead
process. Both high throughput technologies surely miss some
active compounds that are called false negatives. Although
reasonable efforts are done to minimize the rate of false
positive and negative molecules, false positives and
negatives are yet tolerated in both HTS and VS settings [1,
25].
Current computational approaches can manage the ever
increasing amount of data volumes. Recursive partitioning
(RP) is one of the most popular methods when analysing
large data sets [26]. RP separates data sets and builds
decision trees. Along the decision tree separated clusters are
enriched with active molecules. There are several ways to
assess the enrichment, one method of choice is calculating
average values of biological activity at each node. By the end
of the procedure, biological active molecules are
characterized with particular descriptor sets. These descriptor
settings can be applied to search databases for active
compounds [27, 28]. The recently introduced phylogenetic-
like tree algorithm [29], a clustering method that combines
elements of neural nets, genetic algorithms and substructure
analysis, addresses these issues by iterative classification of
compounds and subclustering. In this case, active
compounds can occur in more than one final cluster and can
be associated with several descriptors, thereby providing
diversified structure–activity relationships (SARs).
Labute [30] and coworkers developed a statistical
approach to extract knowledge from HTS experiments,
which is called binary QSAR (bQSAR). bQSAR uses Bayes
theorem to builds models for structural descriptors and
molecular properties to predict biological activity. Xue et al.
achieved 90% accuracy using this method for the prediction
of oestrogen receptor antagonists [31]. Harper et al. [32]
introduced a binary kernel discrimination approach, and
investigated its performance on two datasets. The first is a
set of 1650 monoamine oxidase inhibitors, and the second a
set of 101437 compounds from an in-house enzyme assay.
They compared the performance of binary kernel
discrimination with a simple procedure which they called
“merged similarity search”, and also with a feed forward
neural network. Binary kernel discrimination was shown to
perform robustly with varying quantities of training data and
also in the presence of noisy data. They highlighted the
importance of the judicious use of general pattern
recognition techniques for compound selection [33].
Depending on the size of the data set, rules explaining
biological data can be determined interactively. The method
scales linearly with the number of descriptors, so hundreds
of thousands of structures can be analyzed utilizing
thousands to millions of molecular descriptors. There are
currently no methods to deal with statistical analysis
problems of this size. An important aspect of this analysis is
the ability to deal with mixed datasets. In contrast to most
current quantitative structure/activity relationship (QSAR)
methods requiring compounds with similar binding modes it
could identify SAR rules for classes of compounds in the
same data set that might be binding in different ways [33].
Support vector machines (SVMs) are extensively used
for the interpretation of screening datasets [34]. Zernov et al.
[33] applied SVM to real-life large-scale drug discovery
problems, specifically, the creation of drug- and agro-
Integration of Virtual and High Throughput Screening Combinatorial Chemistry & High Throughput Screening, 2011, Vol. 14, No. 10 893
likeness filters for screening large compound collections.
One particular objective was to compare the performance of
SVM and artificial neural network (ANN) [35] classifiers on
the same data. It was possible not only to define the main
characteristics of derived models but also to reveal a
peculiarity relating to the usage of descriptors and definition
of the border between active and inactive samples. This
peculiarity is critical for deriving adequate classification
models.
COMPLEMENTARITIES IN SCREENING EFFORTS
Considering the primary objective of both virtual and
high throughput screening there are at least three different
strategies of integration including parallel, sequential and
focused screening.
Parallel screening. The parallel approach realizes the
simultaneous application of both VS and HTS tools on the
very same screening library (Fig. 3).
VS HTS
Screening library
VS hits HTS hits
Analysis
Hits
Fig. (3). Parallel screening strategy.
Parallel HTS and structure-based VS study performed on
tyrosine phosphatase-1B inhibitors (PTP1B) – a potential
type II diabetes target - identified separate sets of hits with
low to mid activity ranges [36]. A corporate sublibrary of
approximately 400000 compounds was screened by HTS for
compounds that inhibited PTP1B. Concurrently, molecular
docking was used to screen approximately 235000
commercially available compounds against the X-ray
crystallographic structure of PTP1B, and 365 high-scoring
molecules were tested as inhibitors of the enzyme. Of
approximately 400000 molecules tested in the high-
throughput experimental assay, 85 (0.021%) inhibited the
enzyme with IC50 values less than 100 μM; the most active
had an IC50 value of 4.2 μM. Of the 365 molecules suggested
by molecular docking, 127 (34.8%) inhibited PTP1B with
IC50 values less than 100 μM; the most active of these had an
IC50 of 1.7 μM. Considering only wet tested compounds
structure-based docking therefore enriched the hit rate by
1700-fold over random screening. The hits from both the
high throughput and virtual screens were dissimilar from
phosphotyrosine, the canonical substrate group for PTP1B
and furthermore the two hit lists were very different from
each other. Surprisingly, the docking hits were judged to be
more druglike than the HTS hits. The diversity of both hit
lists and their dissimilarity from each other underlines that
docking and HTS may be complementary techniques for lead
discovery.
Analysis of chemical clusters identified in VS and HTS
campaigns against glycogen synthase kinase-3
(GSK-3
)
also underpinned the complementarity of the methods [37].
GSK-3
is a serine/threonine kinase that has recently
emerged as a key target for neurodegenerative diseases and
diabetes. As an initial step a virtual screen was developed to
discriminate known GSK-3
inhibitors and inactive
compounds using FlexX, FlexX-Pharm, and FlexE. The
maximal enrichment factor (EF = 28) suggested that the
protocol identified potential GSK-3
inhibitors effectively
from large compound collections. The effectiveness of the
screening protocol was further investigated by comparative
HTS and VS performed for the same subset of our corporate
library. Enrichment factors, the significantly higher hit rate
of virtual screening (12.9%) than that of the HTS (0.55%),
and also the comparison of active clusters suggest that the
virtual screening protocol is an effective tool in GSK-3
-
based library focusing. Head-to-head comparison of
true/false positives and negatives revealed the two
approaches to be complementary rather than competitive.
Quantitative HTS and structure-based docking of over
70563 compounds were performed against the antibiotic
resistance target
-lactamase, which has been extensively
characterized enzymologically, crystallographically, and for
some classes of false-positives in screening. The mechanism
of every qHTS active was established, allowing to determine
which mechanisms were most responsible for artefact hits
and how often putatively reactive functional groups were
problematic. Of the 1274 initial inhibitors, 95% were
detergent-sensitive and were classified as aggregators.
Among the 70 remaining were 25 potent, covalent-acting
lactams. Mass spectra, counter-screens, and crystallography
identified 12 as promiscuous covalent inhibitors. The
remaining 33 were either aggregators or irreproducible. No
specific reversible inhibitors were found. Molecular docking
to prioritize molecules from the same library for testing at
higher concentrations was applied afterwards. Of 16 tested, 2
were modest inhibitors. Subsequent X-ray structures
corresponded to the docking prediction. These results
suggest that when attempting to identify small molecule
modulators for a currently “undruggable” or “genomic”
target for which structural information is available, it may be
useful to screen and dock in parallel, using one to help
interpret and guide the other [38].
Shoichet et al. undertook a parallel docking and HTS
screen of 197861 compounds against cruzain, a thiol
protease target for Chagas disease, looking for reversible,
competitive inhibitors. On workup, 99% of the hits were
eliminated as false positives, yielding 146 well-behaved,
competitive ligands. These fell into five chemotypes: two
were prioritized by scoring among the top 0.1% of the
docking-ranked library, two were prioritized by behaviour in
the HTS and by clustering, and one chemotype was
894 Combinatorial Chemistry & High Throughput Screening, 2011, Vol. 14, No. 10
prioritized by both approaches. Determination of an
inhibitor/cruzain crystal structure and comparison of the
high-scoring docking hits to experiment illuminated the
origins of docking false-negatives and false-positives.
Prioritizing molecules that are both predicted by docking and
are HTS-active yields well-behaved molecules, relatively
unobscured by the false-positives to which both techniques
are individually prone [39].
Focused screening. In general CADD approaches are
sufficient enough to filter out molecules from screening
databases which are incompatible with a particular active site
or which are not similar enough to known active compounds.
Moreover, VS strategies are also capable of prefiltering
subsets of screening databases that are more likely to bind to
a given active site. Altogether false and true hits provided by
VS leads to the identification of subsets at a manageable size
of large databases. Focused screening strategies enable the
use of complex low-throughput screening assays like cell-
based high content screening (HCS) assays. However, in
practice a complex compound registration and managing
system is needed to execute a focused screening approach
because hits are selected and identified individually from
plates. Focused screening strategies were shown to be the
most frequently used combined methods over the past few
years (Fig. 4).
VS HTS
Screening library Focused library Hits
Hypothesis generation
Fig. (4). Focused and sequential screening.
Most of the success stories reported in the virtual
screening literature [1] follow this strategy when screening
hundreds of thousands of compounds virtually before testing
the top few dozen experimentally. Since this is the most
popular and extensively reviewed integration strategy we
discuss here only two case studies; one for ligand-based and
another for structure based focusing.
One example of focusing by ligand-based 3D
pharmacophore searching has been used to identify VLA-4
integrin antagonists with submicromolar potency [40]. The
antigen 41 (very late antigen-4, VLA-4) plays an
important role in the migration of white blood cells to sites
of inflammation. It has been implicated in the pathology of a
variety of diseases including asthma, multiple sclerosis, and
rheumatoid arthritis. A series of potent inhibitors of 41,
that were discovered using computational screening for
replacements of the peptide region of an existing tetrapeptide-
based 41 inhibitor (4-[N‘-(2-methylphenyl)ureido] phenyl-
acetyl-Leu-Asp-Val) derived from fibronectin. The search
query was constructed using a model based on the X-ray
conformation of the related integrin-binding region of
vascular cell adhesion molecule-1 (VCAM-1). The 3D
search query consisted of the N-terminal cap and the
carboxyl side chain of the ligand because, upon the basis of
Polgár and Keser
existing structure
activity data on this series, these were
known to be critical for high-affinity binding to 41. The
computational screen identified 12 compounds from a virtual
library of 8624 molecules as satisfying the model and
synthetic filters. All of the synthesized compounds tested
inhibit 41 association with VCAM-1, with the most
potent compound having an IC50 of 1 nM, comparable to the
query compound. A 3D QSAR was generated that
rationalizes the variation in activities of these 41
antagonists. These results demonstrated that it is possible to
rapidly identify nonpeptidic replacements of peptide
antagonists. This scaffold hopping approach could be useful
in identification of nonpeptidic inhibitors with improved
pharmacokinetic properties relative to their peptidic
counterparts.
A structure-based virtual screening was conducted on a
ligand-supported homology model of the human histamine
H4 receptor (hH4R). More than 8.7 million 3D structures
derived from different vendor databases were investigated by
docking to the hH4R binding site using FlexX. A total of 255
selected compounds were tested by radioligand binding
assay and 16 of them possessed significant [3H]histamine
displacement. Several novel scaffolds were identified that
can be used to develop selective H4 ligands in the future.
The average hit rate of the hH4R in vitro screening was
6.3%. The hit rates, concerning compounds selected from
“top_2000”, “top_45000”, and “top_45000_analog” sets
were 4.6, 7.2, and 4.3%, respectively. It is interesting to note
that “top_45000” selection yielded the highest hit rate.
Moreover, the two most potent compounds were also found
by this selection method. These results underline the impact
of hit selection approaches in virtual screening.
“Top_45000_analog” selection that included structural
analogues of the compounds selected from the top 45000
compounds resulted in the smallest hit rate as expected.
Although structural cores were identical to those of the
original hits altered substitution patterns are probably
responsible for differential binding characteristics [41]. This
study also serves the importance of the integration of these
methods.
Sequential screening. The combination of VS-based or
other filter -based subset selection and HTS in an iterative
way is called sequential screening that demonstrates a real
integrated screening strategy (Fig. 4). Differential sequential
strategies can be assessed based on the information
available. When active molecules are known, similarity
search methods can be applied to enrich screening library
subsets with molecules having the desired biological activity.
If resulting subsets contains hit molecules with acceptable
biological activity the subsequent steps will provide
molecules with a better biological activity at a higher
probability. Several active molecules being available as
starting points for the subsequent searching steps can
decrease the size of the screening library because frequency
values can be assigned to candidate molecules based on their
occurrence in the different search steps. In an alternative
way, screening databases can be clustered including or
excluding the available information about the hits. When
known ligands are used those clusters are tested in which
active molecules are found. New hits can then be re-used as
queries in the next clustering step. If there are no known
ligands representative molecules from each cluster can be
Integration of Virtual and High Throughput Screening
selected for biological testing. If these clusters includes
novel hits these can be re-used as query molecules and this
process is repeated until necessary. This iterative approach
has the ability to identify analogues at the very early steps
and provide an initial structure activity relationship that can
facilitate medicinal chemistry efforts. Based on the
sequential screening strategy cluster analysis and data
management IT systems were developed [42]. Sequential
screening is successful if the applied computational filter
enriches the database with active molecules leading to an
increased hit rate in experimental screening. Sequential-
screening applications have already been applied
successfully on many occasions. Sequential screening based
on RP was successfully used for a set of 14 G-protein-
coupled receptors. Recursive partitioning, was used by
Hertzog et al. [43] to help uncover and understand structure–
activity relations and to help biology and chemistry experts
make better decisions on which compounds to screen next
and better characterize. Fig. (5) shows the sequential-
screening process as applied to 14 G-protein–coupled
receptors. The early generation of SAR rules by sequential
screening can greatly improve the efficiency of the screening
process.
Engels and Venkatarangan used sequential screening in
two iterations that resulted in hit rates of up to 40% [44].
Retrospective studies indicate that, by screening from 10 to
20% of a collection in several assays, one can find from 50
to 80% of the active compounds [45].
Two-stage sequential screening experiments were carried
out based on HTS data for 18 cancer cell lines of the NCI
[46]. In the first step, hierarchical clustering was applied to
analyze the screened compound data set consisting of
approximately 32000 compounds, and a diversity-based
selection was carried out producing a representative subset
containing 3000 molecules. Testing of these compounds
already produced relatively high hit rates between 1% and
2% (dependent on the cell line), thus at least 10-fold higher
than expected from random HTS. In the second round,
hierarchical clustering was used to expand obtained hits in
the remainder of the screening set and activity-based
selections were carried out, yielding between 250 and 500
new candidate compounds per cell line. Assays of these
compounds resulted in hits rates between 33% and 43%, thus
initial compound RP SCAP
set analysis
Fig. (5). Sequential screening strategy applied by Hertzog et al. [43].
Combinatorial Chemistry & High Throughput Screening, 2011, Vol. 14, No. 10 895
achieving an overall very significant improvement over
random HTS.
In an effort to identify novel Janus kinase 3 (JAK3)
inhibitors, Chen et al. adopted a sequential focused screening
approach to search in-house chemical database [47].
Sequential screening is unique in that it conducts screening
in an iterative fashion. Instead of attempting to find all the
desirable hits in a single round of screening, it runs screening
in multiple rounds and selects the screening candidates of
each round based on the new structure-activity data coming
out of the previous round of screening. Chen applied this
strategy into a focused screening project to increase the
structural novelty of screening hits. By biologically testing
only 79 selected compounds, they successfully identified 19
compounds showing IC50 < 20 μM, with four of them in the
nanomolar range. With the advantages of efficiency and
flexibility, this approach may be utilized to identify leads for
other therapeutic targets. By applying a sequential focused
screening approach, they not only successfully identified a
couple of novel scaffolds for JAK3 inhibition but also
achieved these findings with a quite high level of screening
efficiency. As a general strategy for compound screening,
this approach can be applied to any chemical databases,
including virtual and non-proprietary databases, greatly
reducing potential acquisition and synthesis costs. It can be
used in association with various computational and
biological screening techniques with great flexibility,
making it a practical tool for current lead identification
processes.
COMPLEMENTARITIES AT THE VS-HTS INTERFACE
Having discussed a variety of methodological aspects, the
final section describes different types of applications at the
interface of VS and HTS (other than those already mentioned in
the context of specific methods). Hit expansion and progression
to leads is well illustrated by a study combining several
modelling approaches [48]. Starting from a hit displaying a
micromolar activity against the μ-opiate receptor, two
complementary concepts were applied by Poulain and
coworkers for analog design and the results were combined to
produce a lead with sub-nanomolar potency for this receptor
[49]. The first method generated analogs of the original hit as
combinations of topologically similar groups representing
compound
evaluation
SAR rules
compound screen
with the hypothesis
screen
cherry picking
compound
collection
896 Combinatorial Chemistry & High Throughput Screening, 2011, Vol. 14, No. 10
different pharmacophore that were present in this hit, thus
preserving the overall “skeleton” of the compound. The second
strategy aimed at identifying analogues displaying
pharmacophore patterns similar to the active compound without
preserving molecular topology. These parallel approaches
proved to be complementary in SAR elucidation and hit
expansion.
As part of a smart screening campaign to target gene
families, screening candidates for kinase and GPCR targets
were selected using neural network simulations and BCUT
descriptors [50]. Neural network simulations identified over
80% of compounds suitable to target each gene family. In
addition, this strategy allowed the identification of compounds
that were active against a kinase but structurally dissimilar to
kinase ligands used in neural network training sets, thus
demonstrating that these VS methods were capable of detecting
compounds having different scaffolds but similar activity.
Another neural network application involved a ligand based
combinatorial design strategy to generate selective purinergic
receptor (A2) antagonists based on a topological
pharmacophore similarity metric using self-organizing maps
[51]. On average, so designed ligands displayed a three-fold
improvement in their binding constants and 3.5-fold higher
selectivity than the ones in the initial library, thereby illustrating
the value of focusing on small activity-enriched libraries for
screening applications.
An information theory-oriented strategy was established to
interface screening and library design and applied to search for
cyclin-dependent kinase inhibitors [52]. Following this
approach, each molecule is considered a “question” and
encoded as a “molecular signature”, a bit string recording the
presence or absence of a particular pre-defined pharmacophore
feature, and the assay result is the “answer”. Informative design
selects molecules in order to maximize the difference between
present and absent chemical features that are responsible for
important compound characteristics based on assay data. In this
study, informative design using 3D pharmacophores
outperformed compound design using 2D substructures, with an
enrichment factor of 7.7 for pharmacophore-based informative
design versus 1.5 for substructure-based selection.
In an anti-bacterial screen of a combinatorial library
containing more than 10000 compounds, a total of 212 active
molecules were detected [53]. However, application of
hierarchical clustering revealed that these actives populated only
seven distinct structural classes, thus providing a more realistic
measure of true hit rates from HTS, which are usually expected
to be in the range between 0.1% and 1%. Subsequently, 11
different nearest neighbour analyses were carried out (in
descriptor spaces representing database compounds) based on
anti-bacterial hits and in each case, between 35 and 66
compounds were selected for follow-up testing. These assays
based on different nearest neighbour selections produced hit
rates between 3% and 36%, with an average of 14% [54].
PRESENT APPLICATIONS AND FUTURE POSITION
For today it became obvious that HTS-based approaches
themselves will not provide as many new chemical entities
as it was expected. The combination of the existing
technologies, however, might contribute to more
sophisticated and hopefully more successful strategies that
Polgár and Keser
will provide better solutions to decrease high attrition rates
during drug discovery. This contribution underpins the view
that VS and HTS have this potential if integrated. Several
examples collected to this review demonstrate that these are
not competitive but complementary approaches of lead
discovery, and that their integration is likely to become
fruitful providing viable chemical leads.
Chemoinformatics driven methodologies can also be
applied in a similar way for the reduction of experimental
efforts. In spite of the fact that theoretical approaches have
limitations and include several approximations, success
stories mentioned here demonstrate that analysis of VS can
influence the success rate of experimental screening.
Presently, this can be well supported by the results of
sequential screening, though there are only a limited number
of relevant studies. It is expected that the continuous
development of VS and HTS technologies will open several
ways to develop focused and iterative screening
methodologies. Integrated and focused approaches may
contribute more to lead discovery programs in the close
future. Bearing this in mind, developing functional interfaces
between theoretical and experimental screening technologies
appears to be highly validated.
REFERENCES
[1] Polgár, T.; Keser, G.M. Virtual Screening, Encyclopedia of
Pharmaceutical Technology, Pharmaceutical Technology,
Informaworld Plc, 2006.
[2] Bajorath, J. Integration of virtual and high-throughput screening.
Nat. Drug Discov. Rev., 2002, 1, 882-894.
[3] Polgár, T.; Keser, G.M. Structure-based virtual screening.
Frontiers in Drug Design & Discovery: Structure-Based Drug
Design in the 21st Century, 2007, 3(1), 477-502.
[4] 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., 1997, 23(1-3), 3-25.
[5] Hann, M.M.; Leach, A. R.; Harper, G. Molecular Complexity and
Its Impact on the Probability of Finding Leads for Drug Discovery,
Journal of Chemical Information and Computational Science,
2001, 41(3), 856-864.
[6] Teague, S.J.; Davis, A.M.; Leeson, P.D.; Oprea, T.I. The design of
leadlike combinatorial libraries. Angewandte. Chemistry, Int. Ed.,
1999, 38(24), 3743-3748.
[7] Walters, P.; Stahl, M.T., Murcko, M.A. Virtual screening - an
overview. Drug Discov. Today, 1998, 3(4), 160-178.
[8] Strategic Pooling of Compounds for High-Throughput Screening,
Mike Hann, Brian Hudson, Xiao Lewell, Rob Lifely, Luke Miller,
and Nigel Ramsden, Journal of Chemical Information and
Computational Science, 1999, 39(5), 897-902.
[9] Willett, P. Similarity-based virtual screening using 2D fingerprints.
Drug Discov. Today, 2006, 11(23-24), 1046-1053.
[10] Willett, P.; Barnard, J.M.; Downs, G.M. Chemical similarity
searching. Journal of Chemical Information and Computer
Sciences, 1998, 38(6), 983-996.
[11] Cross, S. S. J. Improved FlexX docking using FlexS-determined
base fragment placement. J. Chem. Inf. Model., 2005, 45(4), 993-
1001.
[12] Schellhammer, I.; Rarey, M. FlexX-Scan: Fast structure-based
virtual screening PROTEINS: Struct. Funct. Genet., 2004, 57(3),
504-517.
[13] ROCS, © 1997-2010 OpenEye Scientific Software.
[14] 3D alignment, © 1999-2010 ChemAxon Ltd.
[15] Lyne, P.D. Structure-based virtual screening: an overview. Drug
Discov. Today, 2002, 7(20), 1047-1055.
[16] Ghosh, S.; Nie, A.; An, J.; Huang, Z. Combinatorial chemistry and
molecular diversity, Structure-based virtual screening of chemical
libraries for drug discovery. Curr. Opin. Chem. Biol., 2006, 10(3),
194-202.
Integration of Virtual and High Throughput Screening Combinatorial Chemistry & High Throughput Screening, 2011, Vol. 14, No. 10 897
[17] Olah, M.M.; Bologa, C.G.; Oprea, T.I. Strategies for compound
selection. Curr. Drug Discov. Technol., 2004, 1(3), 211-220.
[18] Charifson, P.S.; Walters, W.P. Filtering databases and chemical
libraries. Mol. Divers., 2000, 5(4), 185-197.
[19] Hann, M.; Hudson, B.; Lewell, X.; Lifely, R.; Miller, L.; Ramsden,
N. Strategic Pooling of Compounds for High-Throughput
Screening. J. Chem. Inf.Comput. Sci., 1999, 39(5), 897-902.
[20] Andrews, P.R.; Craik, D.J.; Martin, J.L. Functional group
contributions to drug-receptor interactions. J. Med. Chem., 1984,
27(12), 1648-1657.
[21] Muegge, I.; Heald, S.L.; Brittelli, D. Simple selection criteria for
drug-like chemical matter. J. Med. Chem., 2001, 44(12), 1841-
1846.
[22] Bleicher, K.H.; Böhm, H.J.; Müller, K.; Alanine, A.I. A guide to
drug discovery: Hit and lead generation: beyond high-throughput
screening. Nat. Rev. Drug Discov., 2003, 2, 369-378.
[23] White, R.E. High-throughput screening in drug metabolism and
pharmacokinetic support of drug discovery. Ann. Rev. Pharmacol.
Toxicol., 2000, 40(133-157), 133-157.
[24] Mayr, M.M. The future of high-throughput screening. J. Biomol.
Screen., 2008, 13(6), 443-448.
[25] Stahura, F.L.; Bajorath, J. Virtual screening methods that
complement HTS. Comb. Chem. High Throughput Screen., 2004,
7(4), 259-269.
[26] Heping, Z.; Burton, H.S. Recursive Partitioning and Applications.
2nd ed., Springer, 2010.
[27] Chen, X.; Rusinko, A; Young, S.S. Recursive partitioning analysis
of a large structureactivity data set using three-dimensional
descriptors. J. Chem. Inf. Comp. Sci., 1998, 38(6), 1054-1062.
[28] Rusinko, A.; Farmen, M.W.; Lambert, C.G.; Brown, P.L.; Young,
S.S. Analysis of a large structure/biological activity data set using
recursive partitioning. J. Chem. Inf. Comp. Sci., 1999, 39(6), 1017-
1026.
[29] Tamura, S.Y.; Bacha, P.A.; Gruver, H.S.; Nutt, R.F. Data analysis
of high-throughput screening results: Application of multidomain
clustering to the NCI anti-HIV data set. J. Med. Chem., 2002,
45(14), 3082-3093.
[30] Labute, P. Binary QSAR: a new method for the determination of
quantitative structure activity relationships. Pac. Symp. Biocomput.,
1999, 4, 444-455.
[31] Xue, L.; Godden, J.; Gao, H.; Bajorath, J. Identification of a
preferred set of molecular descriptors for compound classification
based on principal component analysis. J. Chem. Inf. Comput. Sci.,
1999, 39(4), 699-704.
[32] Harper, G.; Bradshaw, J.; Gittins, J.C.; Green, D.V.S.; Leach, A.R.
Prediction of biological activity for high-throughput screening
using binary kernel discrimination. J. Chem. Inf. Comput. Sci.,
2001, 41(5), 1295-1300.
[33] Zernov, V.Z.; Balakin, K.V.; Ivaschenko, A.A.; Savchuk, N.P.;
Pletnev, I.P. Drug discovery using support vector machines. The
case studies of drug-likeness, agrochemical-likeness, and enzyme
inhibition predictions. J. Chem. Inf. Comput. Sci., 2003, 43(6),
2048-2056.
[34] Ivanciuc, O. Applications of support vector machines in chemistry.
Rev. Comput. Chem., 2007, 23, 291-400.
[35] Cheng, B.; Titterington, D.M. Neural networks: A review from a
statistical perspective. Stat. Sci., 1994, 9(1), 2-30.
[36] Doman, T.N.; Mcgovern, S.L.; Witherbee, B.J.; Kasten, T.P.;
Kurumbail, R.; Stallings, W.C.; Connolly, D.T.; Shoichet, B.K.
Molecular docking and high-throughput screening for novel
inhibitors of protein tyrosine phosphatase-1B. J. Med. Chem., 2002,
45(11), 2213-2221.
[37] Polgar, T.; Baki, A.; Szendrei, G.I.; Keseru G.M. Comparative
virtual and experimental high-throughput screening for glycogen
synthase kinase-3 inhibitors. J. Med. Chem., 2005, 48(25), 7946-
7959.
Received: September 3, 2010 Revised: February 4, 2011
[38] Babaoglu, K.; Simeonov, A.; Irwin, J.J.; Nelson, M.E.; Feng, B.;
Thomas, C.J.; Cancian, L.; Costi, M.P.; Maltby, D.A.; Jadhav, A.;
Inglese, J.; Austin, C.P.; Shoichet, B.K. Comprehensive
Mechanistic Analysis of Hits from High-Throughput and Docking
Screens against -Lactamase. Journal of Medicinal Chemistry, 2008,
51(8), 2502-2511.
[39] Ferreira, R.S.; Simeonov, A.; Jadhav, A.; Eidam, O.; Mott, B.T.;
Keiser, M.J.; McKerrow, J.H.; Maloney, D.J.; Irwin, J.J.; K.
Shoichet, B.K. Complementarity between a docking and a high-
throughput screen in discovering new Cruzain inhibitors. J. Med.
Chem., 2010, 53(13), 4891-4905.
[40] Singh, J.; van Vlijmen, H.; Liao, Y.; Lee, W-C.; Cornebise, M.;
Harris, M.; Shu, I.; Gill, A., Cuervo, J.H.; Abraham, W.M.; Adams,
S.P. Identification of potent and novel 41 antagonists using in
silico screening. J. Med. Chem. 2002, 45(14), 2988-2993.
[41] Kiss, R.; Kiss, B.; Könczöl, Á.; Szalai, F.; Jelinek, I.; László, V.;
Noszál, B.; Falus, A.; Keser, G.M. Discovery of novel human
histamine H4 receptor ligands by large-scale structure-based virtual
screening. J. Med. Chem., 2008, 51(11), 4891-4905.
[42] Stahura, F.L.; Xue, L.; Godden, J.W.; Bajorath, J. Methods for
compound selection focused on hits and application in drug
discovery. J. Mol. Graph. Model., 2002, 20(6), 439-446.
[43] Jones-Hertzog, D.K.; Mukhopadhyay, P.; Keefer, C.E.; Young,
S.S. Use of recursive partitioning in the sequential screening of G-
protein – coupled receptors. J. Pharmacol. Toxicol. Methods, 2008,
51(2), 3145-3153.
[44] Engels, M.F.M.; Venkatarangan, P. Smart screening: Approaches
to efficient HTS. Curr. Opin. Drug Discov. Devel., 2001, 4, 275-
283.
[45] 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. Infor.
Comp. Sci., 1998, 38(3), 511-522.
[46] Stanton, D.; Morris, T.W.; Roychoudhury, S.; Parker, C.N.
Application of nearest-neighbour and cluster analyses in
pharmaceutical lead discovery. J. Chem. Inform. Comp. Sci., 1999,
39(1), 21-27.
[47] Chen, X.; Wilson, L.J.; Malaviya, R.; Argentieri, R.L., Yang, S.M.
Virtual screening to successfully identify novel Janus kinase 3
inhibitors: A sequential focused screening approach. J. Med.
Chem., 2008, 51, 7015-7019.
[48] Good, A.C.; Krystek, S.R.; Mason, J.S. High-throughput and
virtual screening: core lead discovery technologies move towards
integration. Drug Discov. Today, 2000, 5(12), 61-69.
[49] Poulain, R.F.; Horvath, D.; Bonnet, B.; Eckhoff, C.; Chapelain, B.;
Bodinier, M-C.; Deprez, B. From hit to lead. Analyzing structure-
profile relationships J. Med. Chem., 2001, 44(21), 3391-3401.
[50] Manallack, D.T.; Pitt, W.R.; Gancia, E.; Montana, J.G.;
Livingstone, D.J.; Ford, M.G.; Whitley, D.C. Selecting screening
candidates for kinase and G protein-coupled receptor targets using
neural networks. J. Chem. Inf. Comput. Sci., 2002, 42(5), 1256-62.
[51] Schneider, G.; Nettekoven, M. Ligand-based combinatorial design
of selective purinergic receptor (A2A) antagonists using self-
organizing maps. J. Comb. Chem., 2003, 5(3), 233-237.
[52] Bradley, E.K.; Miller, J.L.; Saiah, E.; Grootenhuis, D.J.
Informative library design as an efficient strategy to identify and
optimize leads: Application to cyclin-dependent kinase 2
antagonists. J. Med. Chem., 2003, 46(20), 4360-4364.
[53] Grüneberg, S.; Stubbs, M.T.; Klebe, G. Successful virtual
screening for novel inhibitors of human carbonic anhydrase:
Strategy and experimental confirmation. J. Med. Chem., 2002,
45(17), 3588-3602.
[54] Spencer, R.W. Diversity analysis in high throughput screening. J.
Biomol. Screen, 1997, 2(2), 69-70.
Accepted: June 11, 2011