Current Topics in Medicinal Chemistry, 2011, 11, 000-000 1
1568-0266/11 $58.00+.00 © 2011 Bentham Science Publishers Ltd.
Chemogenomics Approaches for Receptor Deorphanization and Exten-
sions of the Chemogenomics Concept to Phenotypic Space
Eelke van der Horst1, Julio E. Peironcely2,3, Gerard J. P. van Westen1, Olaf O. van den Hoven1,
Warren R. J. D. Galloway4, David R. Spring4, Joerg K. Wegner5, Herman W. T. van Vlijmen5,
Ad P. IJzerman1, John P. Overington6 and Andreas Bender7,*
1Division of Medicinal Chemistry, Leiden / Amsterdam Center for Drug Research, Einsteinweg 55, 2333 CC Leiden, The
Netherlands; 2Division of Analytical BioSciences, Leiden / Amsterdam Center for Drug Research, Einsteinweg 55, 2333
CC Leiden, The Netherlands; 3TNO, Quality of Life, Utrechtseweg 48, 3704 HE Zeist, The Netherlands; 4Department of
Chemistry, University of Cambridge, Lensfield Road, Cambridge CB2 1EW, United Kingdom; 5Tibotec BVBA, Generaal
De Wittelaan L 11B 3, 2800 Mechelen, Belgium; 6EMBL Outstation – Hinxton, European Bioinformatics Institute, Well-
come Trust Genome Campus, Hinxton, Cambridge CB10 1SD, United Kingdom; 7Unilever Centre for Molecular Infor-
matics, Department of Chemistry, University of Cambridge, Lensfield Road, Cambridge CB2 1EW, United Kingdom
Abstract: Chemogenomic approaches, which link ligand chemistry to bioactivity against targets (and, by extension, to
phenotypes) are becoming more and more important due to the increasing number of bioactivity data available both in
proprietary databases as well as in the public domain. In this article we review chemogenomics approaches applied in four
different domains: Firstly, due to the relationship between protein targets from which an approximate relation between
their respective bioactive ligands can be inferred, we investigate the extent to which chemogenomics approaches can be
applied to receptor deorphanization. In this case it was found that by using knowledge about active compounds of related
proteins, in 93% of all cases enrichment better than random could be obtained. Secondly, we analyze different chemin-
formatics analysis methods with respect to their behavior in chemogenomics studies, such as subgraph mining and Baye-
sian models. Thirdly, we illustrate how chemogenomics, in its particular flavor of ‘proteochemometrics’, can be applied to
extrapolate bioactivity predictions from given data points to related targets. Finally, we extend the concept of ‘chemoge-
nomics’ approaches, relating ligand chemistry to bioactivity against related targets, into phenotypic space which then falls
into the area of ‘chemical genomics’ and ‘chemical genetics’; given that this is very often the desired endpoint of ap-
proaches in not only the pharmaceutical industry, but also in academic probe discovery, this is often the endpoint the ex-
perimental scientist is most interested in.
Keywords: Chemogenomics, proteochemometrics, deorphanization, GPCR, virtual screening, G-protein coupled receptors,
orphan receptors, target prediction, mode of action analysis.
INTRODUCTION
The term ‘chemogenomics’[1], first coined in 2001[2,3]
represents the systematic study of ligand chemistry and pro-
tein targets (or generally gene products) they show bioactiv-
ity against [4]. While in this kind of study the links between
the underlying structure connecting ligand chemical space
and biological bioactivity space are of primary interest, the
related terms ‘chemical genetics’ and ‘chemical genom-
ics’[5,6] focus more on the ability of small molecule chemis-
try to modulate biological systems in a specific and directed
manner, similar to classical genetic approaches such as
knock-out organisms. However, chemical tools can act as
both protein activators and inhibitors, and they can be ap-
plied at various developmental stages and different concen-
trations; hence, they are often more flexible to use than their
biological counterparts.


*Address correspondence to this author at the Unilever Centre for Molecular
Informatics, Department of Chemistry, University of Cambridge, Lensfield
Road, Cambridge CB2 1EW, United Kingdom; Tel: +44 (1223) 762 983;
Fax: ?????????????; E-mail: andreas.bender@cantab.net
In recent years more and more knowledge about the
chemistry of bioactive molecules has entered the public do-
main, in databases such as DrugBank [7], BindingDB [8],
PDSP Ki [9] and so on, as well as (partially) cellular assay
databases like PubChem Bioassay [10] and Chembank
[11](see recent reviews such as [12] and [13] for a more
comprehensive overview of these). Most recently, and as one
of the biggest efforts of its kind, the European Bioinformat-
ics Institute (EMBL-EBI) via a grant from the Wellcome
Trust acquired the rights of the StARlite database previously
sold by Inpharmatica, which added in excess of five hundred
thousand bioactive chemicals with millions of binding and
functional assay data points to the publicly available bioac-
tivity knowledge in the form of the ChEMBL database [14]
(discussed in a recent review [15]). Apart from databases
linking chemical structures to protein targets, also pheno-
typic databases such as those capturing side effects of drugs
are now becoming publicly available[16], increasing the
amount of data available for chemogenomics studies that
does not stop at the biochemical level but also extends this
information to the organism level effects of chemical struc-
tures.

2 Current Topics in Medicinal Chemistry, 2011, Vol. 11, No. 16 van der Horst et al.
What is interesting to note is that publications containing
the terms ‘chemogenomic’ or ‘chemogenomics’ in the Topic
(as defined by Web of Science; containing publication title,
abstract and keywords) has not seen the exponential growth
one might have expected in recent years. This trend is visual-
ized in Fig. (1); while citation data for 2010 are incomplete,
it seems as if from 2005 to 2010 an approximate plateau
phase of only around 25 publications per year were achieved.
In absolute terms this is a relatively small number, and it
would be an interesting subject of discussion what the reason
would be – whether the concept did not work out in practice;
whether people already switched to using other terms instead
of these two; or whether the real boom of the area is still
ahead of us. The future will certainly tell.










Fig. (1). Citations of ‘chemogenomics OR chemogenomic’ in Topic
(Title, Abstract, Keywords) of Web of Science (as on 1 June 2010).
While citation data for 2010 are incomplete, it seems as if from
2005 to 2010 an approximate plateau phase of around 25 publica-
tions per year were achieved.

Large, well indexed ligand-target-assay databases are at
the very heart of chemogenomics studies which aim to relate
ligand chemistry to the similarity of protein targets on a large
scale; and it should be kept in mind that data completeness
determines the analysis outcome [17,18]. Despite informa-
tion not being perfect, it can still be of value in a more in-
formed decision-making process. It was noted in pharmaceu-
tical companies that information from related targets can be
used to steer ligand discovery for an orphan target of interest
in a productive direction, but at the same time it was by no
means obvious in which way, and in which cases, this could
be achieved. In parallel, the previously eminent paradigm of
‘ligand selectivity’ has recently been replaced in favor of a
desired polypharmacology profile [19-21].Such a profile
requires not only a set of targets of interest from the biologi-
cal side that needs to be considered, but also ligand chemis-
try to be chosen so they can be bioactive against multiple
targets in parallel. However, important problems are still
unsolved for this paradigm, for example the question how a
researcher defines a desired bioactivity in the first place; and
how bioactivity profiles actually translate from animal sys-
tems to humans. This controlled polypharmacology is proba-
bly more easily achieved in some cases (e.g. a set of struc-
turally related Class A GPCRs) than in case of others (e.g.
inhibiting a protease and antagonizing a GPCR simultane-
ously). A recent study by one of the authors [21] was analyz-
ing bioactive chemical space, as defined by the WOMBAT
database, ECFP4 circular fingerprints and a Principal Com-
ponent Analysis of the resulting Bayes Models per class, the
result of which is visualized in Fig. (2). This figure can be
interpreted as follows: PC1 and PC2 are the axes of maxi-
mum chemical ligand diversity; hence, the classes with the
highest loadings along both axes have the most different
chemistry in the set, here dopamine D2 receptor ligands and
HIV integrase inhibitors. Along PC3
opioid receptor
ligands possess the furthest distance from those two classes.
And indeed, when analyzing ligand chemistry in more detail
(see [21] for further details) we can see that D2 ligands al-
ways possess a tertiary nitrogen and often unsaturated six-
membered rings; HIV integrase inhibitors typically possess
catechol moieties and carboxylic acid groups, frequently
they also are esters; and
opioid receptor ligands show con-
siderable diversity, from the complex morphine scaffold to
much more diverse ligands that could also resemble enzyme
inhibitors and GPCR ligands. Similar analyses, based some-
times on different datasets and different algorithms, were
also performed by other groups such as at Pfizer[22,23],
UCSF[24] and the University of Barcelona [25].
One concept that is implicit in chemogenomics ap-
proaches is that chemical space is reasonably well behaved,
continuous and can be interpolated. With respect to the bio-
logical side, it is assumed that ligands active against one
protein are more often than random also active against a re-
lated protein structure. Very much related to these concepts
is the idea of ‘affinity fingerprints’ published by Kauvar et
al. [26]. In its simplest form, affinity fingerprints make the
assumption that affinity to novel targets can be approximated
from known affinities to a set of (‘orthogonal’) proteins.
Later this idea was extended to the computational domain
with concepts such as docking-based fingerprints[27] and
ligand based ‘Bayes Affinity Fingerprints’[21], but what is
interesting in the chemogenomics context is the following.
As current chemogenomics thinking goes, only measures of
protein similarity are related to measures of ligand similarity
- and, hence, simply ligands of related receptors to an orphan
receptor are used as a starting point for de-orphanization
projects. However, this is only the simplest version of the
affinity fingerprints imaginable. According to the affinity
fingerprint concept, also bioactivity values against very dif-
ferent proteins could be used to make (approximate) predic-
tions as to which ligands are active against a currently or-
phan receptor in a linear combination of affinities, or a yet
unknown more complex function. Given that more and more
data becomes available it is surely only a matter of time until
the current chemogenomics concepts of ‘similar targets bind
similar ligands’ get extended to also cover more complex
relationships between ligand chemistry and target bioactivity
space. We anticipate that research in this area will have huge
benefits for experimental biologists as well as biological and
medicinal chemist and will be transformative for the com-
mercial life-science sector.
In a similar vein, the term proteochemometrics is cer-
tainly very dissimilar at first sight to chemogenomics; how-
ever, the underlying ideas of both concepts are not funda-
mentally different. Both of them attempt to relate ligand
chemistry to a set of different proteins instead of a single
protein (which would be more the domain of conventional

Chemogenomics for Deorphanization and Phenotype Linkage Current Topics in Medicinal Chemistry, 2011, Vol. 11, No. 16 3
structure-activity modeling), and the difference between both
concepts are sometimes minor, with proteochemometric
modeling being often more focused on a more defined set of
targets with the aim to model bioactivity relationships quan-
titatively; however none of these characteristics are true in
every case. One of the most comprehensive recent studies
[28] was attempting to model all protein-ligand interaction
space of enzymes, as taken from protein-ligand cocrystals. A
model was generated on 826 pairs of proteins, druglike
ligands and binding affinity or dissociation constant of the
particular protein-ligand pair from AffinDB, PDB Bind,
Binding MOAD and Protein-Ligand Database, and 542 en-
tries from Brenda were used as a test set. Predictions for the
external test set achieved an r2 of 0.53 and a RMSEP of 1.5
(over a pKi range reported[28] to be from 0.7-11.0); hence,
even when using very diverse protein data in this kind of
modeling exercise affinity data can be modeled relatively
reliably – which empirically underlines the validity of the
concepts behind chemogenomic and proteochemometric
modeling.
In this publication no comprehensive overview of all
Chemogenomics studies can be given, and hence a set of
recent reviews in the area should be mentioned [29-36]. In
the following, we will now outline our recent advances in
chemogenomics studies applied to receptor deorphanization,
different representations of molecules in studies of this type,
proteochemometrics studies performed in our group, as well
as extensions of the chemogenomics concept into phenotype
space.
CHEMOGENOMICS APPLIED TO RECEPTOR DEO-
RPHANIZATION
As outlined in the introduction, the most straightforward
analysis in the chemogenomics spirit (which is still depend-
ent on quite a lot of variables such as the precise dataset cho-
sen, the chemical representation used, and the distance met-
ric employed) is to relate ligand similarity to protein similar-
ity in a given bioactivity data set. Previous studies in the
field exist, very early in the field of kinases [37] where struc-
tural binding site similarity was compared to ligand SAR;
however only 58 ATP-site ligands were contained in the
PDB at the time of the article. More recently this database
has been extended significantly, with large-scale profiling
data of kinases entering the public domain [38, 39]. Data of
this type has also been used recently to construct a ligand-
based kinase tree [40] as well as to enzyme families [41].
In the work performed in our group though GPCRs are
the main target family of interest, and less chemogenomics
work has yet been published in this area. One recent study
[42] attempts indeed a (hypothetical) deorphanization of
GPCRs, leaving all ligands of the receptor under considera-
tion out of the training set and generating models using dif-
ferent kernel-based approaches. It was found that using a
binding-pocket kernel that only takes the sequence of resi-
dues in the binding pocket into account, achieve 78.1% cor-
rect predictions, averaged over all classes, on the dataset,
compared to 50% in case of random class assignments.
However, what was not done was the analysis in which cases
chemogenomics studies applied to GPCR ligands would suc-
ceed in practice, and to ‘re-draw’ the GPCR tree based on a
chemogenomics analysis; this is precisely what we per-
formed in our ongoing work in the group that is currently
being validated prospectively.
Other related work has been published by Bock et al.
[43] and Weill and Rognan [44]. Bock et al. [43] described
receptors by using numerical values for the surface tension,
isoelectric point, and accessible surface area for each amino
acid of the receptor as well as a connectivity matrix, ex-
tended by atomic properties such as ionization energy, elec-
tron affinity and atom density on the ligand side and found
that out of about 2,000,000 novel compounds only 2% are
predicted to be active on a novel, orphan receptor. Models
















Fig. (2). Two first principal components of ligand bioactivity space. Inhibitors of HIV-1 integrase and ligands of the dopamine D2 receptor
define the most distinct ligand classes from the chemical side. (Reprinted with permission from [21]).

4 Current Topics in Medicinal Chemistry, 2011, Vol. 11, No. 16 van der Horst et al.
were generated by a Support Vector Machine on 5,319 re-
ceptor-ligand pairs from the PDSP database; however, no
statistics for the models generated has been provided for a
target deorphanization exercise.
In Weill and Rognan [44], a novel protein-ligand finger-
print, termed PLFP, was introduced that captures pharma-
cophoric properties of the ligands, as well as of the trans-
membrane ligand binding pocket. SHED, topological auto-
correlation descriptors (DistFP) and MACCS keys were
combined with SVM, Naïve Bayes and Random Forests to
evaluate the influence of different chemical representations
and model generation methods on predictive performance
and models were evaluated on two external test sets contain-
ing a total of 60 data points. It was found that picking
ligands for targets was usually easier to achieve than the re-
verse, predicting targets for ligands. Also, both global mod-
els, considering all GPCR ligand data in a single model, as
well as a set of 19 local models, one for each subclass of
GPCRs, were compared. Here it was found that DistFP with
support vector machines on local models outperforms on
average all other model generation methods – a hint that lo-
cal bioactivity models may be required to model larger areas
of bioactivity space.
We now extended the above study, and analyzed chemi-
cally in more detail in which cases a chemogenomics classi-
fication of GPCR ligands would be more likely to be suc-
cessful in a deorphanization study than in others. Given the
dependence of this type of study on a large number of factors
(the diversity of chemical and biological space covered in the
datasets; the measures used to establish similarities in both
spaces; and the decision which data points of ‘neighboring’
sequences to include, just to name a few) we felt the need to
establish some practical guidelines for future work in this
area.
For this purpose, we used a receptor-based phylogenetic
classification and compared it to a ligand-based classifica-
tion that is based on exhaustive substructure mining. (Note
that a full primary research article on this work has been
published recently with more details of the results and a fur-
ther interpretation which can be found in this reference [45].)
For the protein side, a set of 44 amino acid residues likely to
be involved in ligand binding of class A GPCRs were se-
lected, based on previous work of Gloriam et al. [46]. This
selection of residues is based on previous work by Surgand
et al. [47] of 30 residues, but extends it by taking the re-
cently published crystal structures of the human
2 and tur-
key
1 adrenoceptors, as well as that of the human adenosine
A2A receptor and their residues involved in ligand binding
into account to extend the previous set to 44 relevant resi-
dues. Based on those amino acid residues, a phylogenetic
tree of class A GPCRs was constructed. An important nu-
ance to classical sequence-based phylogenetic analysis is in
the choice of the scoring matrix for amino-acid interchanges,
since very different effects are anticipated for conserved mo-
lecular recognition effects compared to the different con-
straints observed in protein evolution. For example, glutamic
acid and arginine are generally similar from a protein struc-
ture viewpoint, where they are regularly interchanged on the
surface of proteins; however, when these amino acids are
involved in the binding of a small molecule ligand they have
very different properties. Again, looking to the future, we see
investigating this area as being productive now that signifi-
cant amounts of data are becoming available.
Complementary to these sequence-based classifications a
substructure-based classification of the receptors from the
ligand side was performed in parallel. Exhaustive substruc-
ture mining of GPCR ligands was performed as described in
detail in the primary research article [48]. Structures were
represented as labeled graphs with aromatic bonds being
assigned a different bond type. In this study, the minimum
support value was set to 30% of the number of ligands in
each activity set, meaning that only substructures present in
at least 30% of the ligands were considered in the further
analysis. Substructures below 50 Dalton were discarded,
since very small fragments are chemically not amenable to
interpretation, which is easily the case for larger substruc-
tures of molecules since they are molecular representations
accessible to the way of thinking of a chemist. To calculate
the similarity between activity classes, the Pearson Correla-
tion Coefficient between all feature frequencies in each ac-
tivity class was calculated. The correlation coefficient was
then transformed into a distance measure by subtracting the
correlation coefficient from 1 and using linear scaling of all
results to [0;1]. Tree construction might in principle be influ-
enced by the order in which targets are provided to the tree
constructor, and in order to investigate this effect the target
input order was randomized 10 times and 10 new trees were
generated and compared to each other. Only in very rare
cases the trees generated were different though, supporting
the robustness of the method employed.
The tree that was built based on the multiple sequence
alignment as defined by Gloriam et al. [46] set is shown in
Fig. (3). Four clusters are clearly defined in the tree, namely
the aminergic receptors, adenosine receptors, prostanoid re-
ceptors, as well as the peptide-binding receptors. This clus-
tering is very close to the ‘conventional’ results obtained
from full-sequence phylogenetic trees; note that minor dif-
ferences to the original tree presented by Gloriam et al. are
visible since our target-based tree was only constructed using
bioactivity classes for which sufficient ligand data was avail-
able. For comparison, the ligand-based receptor classifica-
tion tree is provided in Fig. (4). Overall, to a large extent the
target-based phylogenetic tree is conserved, with differences
only visible for some of the receptor subclasses. It should be
kept in mind that this may also partly be due to the chemistry
tested against particular receptors – most scientists would
test known ligands against one receptor subtype also against
another receptor subtype, and hence due to this ‘selection
bias’ some classes might move closer together than with
more even sampling of chemical space screened against a
receptor.
Except for the two purinergic receptors (P2Y1 and
P2Y12) and the two glycoprotein hormone receptors (FSH
and LH), all other receptor pairs represented by two subtype
are clustered together. The adenosine receptors (ADORA1,
ADORA2A, ADORA2B, ADORA3) group together with the
A2B receptor being the most dissimilar from the other,
which is consistent with previous results obtained by our
group. The purinergic receptor P2Y12 is from the ligand-
side found to be similar to the adenosine receptors, which is

Chemogenomics for Deorphanization and Phenotype Linkage Current Topics in Medicinal Chemistry, 2011, Vol. 11, No. 16 5
























Fig. (3). Phylogenetic tree of human class A GPCRs based on the 44 residues identified relevant for binding by Gloriam et al. The color
codes are as follows: black – receptor with aminergic ligands; dark grey – purinergic ligands and melatonin ligands; medium grey – lipid
ligands; light grey – adenosine ligands. (Reprinted with permission from [45]; also see this reference for figure in color.)





























Fig. (4). Phylogenetic tree of class A GPCRs based on frequent substructure mining of GLIDA and ChEMBL data. The color codes are as
follows: black – receptor with aminergic ligands; dark grey - purinergic ligands and melatonin ligands; medium grey – lipid ligands; light
grey – adenosine ligands. (Reprinted with permission from [45]; also see this reference for figure in color.)

6 Current Topics in Medicinal Chemistry, 2011, Vol. 11, No. 16 van der Horst et al.

understandable due to the common purine core typical for
ligands of both subfamilies. The muscarinic acetylcholine
receptors M1, M3, M4, and M5 cluster together as one
group, supporting the low subtype selectivity of muscarinic
antagonists; however, the acetylcholine receptor M2 is found
more distant from this cluster which may be the result of
inclusion of allosteric ligands. Some clusters not present in
the sequence-based tree are present in the ligand-based clas-
sification and they can often be well understood from the
chemical point of view; e.g. the grouping of the eight prosta-
noid receptors displayed in Fig. (3). This cluster is based on
the fact that most prostanoid receptor ligands are direct de-
rivatives of the endogenous ligands, the so-called eicosa-
noids. These ligands form a very homogeneous group which
is dominated by relatively long alkyl chains. The clustering
of the leukotriene and cannabinoid receptors in this lipid
cluster may seem strange at first; however, arachidonic acid
is the common precursor for eicosanoids and two derivatives
of arachidonic acid, anandamide and 2-arachidonyl-glycerol,
both of which are endogenous ligands (‘endocannabinoids’)
of the cannabinoid receptors and which explains the chemi-
cal similarity of ligands in this cluster, and hence, their relat-
edness from the chemical point of view.
While in the original work on this topic [45] the differ-
ences and commonalities between biological and chemical
phylogenetic trees have been analyzed in detail, in the con-
text of this review its applications to the deorphanization of
GPCRs should be commented on in particular. In order to
resemble a real-world setting, we performed a hypothetical
‘de-orphanization exercise’. To do this, we excluded in turn
all ligands of each receptor in the dataset; we so ‘orphanized’
the receptor in this particular analysis run. Next, we ‘de-
orphanized’ the receptor again by predicting its ligands by
using a ligand-based bioactivity model derived from the
closest neighbors of the receptor in sequence space. It was
found that in 93% of the cases our hypothetical de-orphani-
zation exercise was successful, with the model providing
enrichment curves better than random (AUC > 0.5); and for
35% of receptors performance was even ‘good’ (as defined
by PipelinePilot, with an AUC > 0.7). Sample plots for four
receptors are shown in Fig. (5), namely for CHRM1 (mus-
carinic acetylcholine receptor M1), AGTR2 (angiotensin II
receptor, type 2), P2RY1 (purinergic receptor P2Y, G-
protein coupled, 1) and BRS3 (bombesin-like receptor 3).
We could indeed find a rationale in which de-orphanization
exercises are more likely to succeed than in others: The poor
performance concerning the P2RY1 receptor is probably due
to the nature of its ligands, since this set consists of a small
number of highly similar ligands that all possess at least one
phosphate group, a feature not found in other ligands in the
database. (In fact, its most related sibling from the biological
side, P2RY12 Fig. (3), moves far away in chemical space
Fig. (4), illustrating the dissimilarity of the ligands) – and
hence, in this case the deorphanization exercise fails. On the
other hand, in case of CHRM1 very much related ligands are
present in the dataset for its nearest neighbor, CHRM5 –
hence, in this case the deorphanization exercise succeeds.
Overall in this study our method was relatively successful to
achieve this task since for 93% of the receptors studied per-
formance better than random was achieved (AUC > 0.5), and
for 35% of receptors even models with reasonable quality
were obtained (AUC > 0.7).
CHEMICAL REPRESENTATIONS IN CHEMOGENO-
MICS APPROACHES
As in ligand-based virtual screening [49-51], the question
how to represent chemicals in a chemogenomics study, as
well as how to calculate the similarity or distance between
different compound classes and hereby between receptors, is
by no means obvious. Different groups use different ap-
proaches, from the pairwise comparison of compounds in
each class followed by calculating expectation values [24] to
calculating frequency vectors of circular fingerprints for each
class and calculating the correlation coefficient between
them [49,52]. In our previous work, with the aim to obtain a
non-biased representation of chemical substructures, we used
exhaustive substructure enumeration of databases to repre-
sent chemical structures which is outlined in the following
and described in detail in a recent primary research article
[48].
Molecular subgraph mining (also known as association
learning) has two characteristics that render it very different
from other molecular representations such as fingerprints
and fragment-based representations: On the one hand it is a
very demanding method to represent molecules due to the
number of subgraphs that can be generated even for a drug-
like small molecule; this number is in the order of millions
for example for steroids, due also to the complex ring system
present in them. On the other hand it is an unbiased represen-
tation of molecules (apart from the model assumption of
‘atoms’ and ‘bonds’) – no assumptions are made which mo-
lecular features are relevant for a certain property such as a
bioactivity against a protein. This is the case in fingerprints
(‘circular features’ are important e.g. for circular finger-
prints, or ‘pairs of features are related to the property being
considered’ in case of atom pairs, and so on) as well as key-
based fingerprints, where precise chemical groups are pre-
defined. Hence, in our work on the chemogenomics data
mining of GPCR databases presented in detail recently [48]
subgraph mining was employed to analyze the databases as
hand.
In our particular case, molecular structures are repre-
sented as labeled graphs with only heavy atoms being con-
sidered in the analysis. Four types of chemical representation
were used: the initial chemical structure representation with
the atom and bond types unchanged, and three ‘Elaborate’
Chemical Representations (ECRs) that attach labels to at-
oms, as well as bonds, in order to generalize the chemistry
encountered (similar to the typing of atoms in pharmacopho-
res as lipophilic, charged, hydrogen bond donor/acceptor,
etc.). [53,54] which is illustrated in Fig. (6). Atoms and
bonds can both be typed; in our case aliphatic nitrogen, oxy-
gen, and sulfur atoms were represented as aliphatic heteroa-
tom by replacement with the symbol ‘No’ and an extra label
was attached to nitrogen and oxygen atoms in order to indi-
cate the type and number of bound hydrogen atoms. This
captures information similar to hydrogen bond donor and
acceptor properties of a heteroatom. The halogen atoms, Cl,
Br, I, and F, were replaced by X, since they are partially
negatively charged and, hence, can be isosteres in many

Chemogenomics for Deorphanization and Phenotype Linkage Current Topics in Medicinal Chemistry, 2011, Vol. 11, No. 16 7
















Fig. (5). Deorphanization enrichment for different sample bioactivity classes (for full receptor names see main text). Overall, for 93% of the
receptors studied performance better than random was achieved (AUC > 0.5), while for 35% of receptors good-quality models were obtained
(with an AUC > 0.7). (Reprinted with permission from [45])













Fig. (6). A sample molecule in normal (I) and three elaborate
chemical representations, the first one including aromatic bonds
(II), a second one including both aromatic atoms and bonds (III),
and one capturing planar ring systems (IV). In the normal represen-
tation (I), aromatic bonds are represented as alternating single and
double bonds, while in the first elaborate representation (II), a spe-
cial aromatic bond type exists (representation (III) extends aro-
matic typing also to atoms). In both elaborate chemical representa-
tions, wildcards are used for heteroatoms (‘No’) and for halogens
(‘X’). (Reprinted with permission from [48])

(though, in particular in case of fluorine, not all) biological
situations. Also one elaborate representation includes a spe-
cial bond type for aromatic bonds, while the second repre-
sentation also has a special type for aromatic atoms. Finally,
the third representation offers a special type for planar ring
systems, which has been successfully applied previously to
predict the mutagenicity of compounds[54]. Overall, it was
found that the elaborate representations indeed outperform
the original representation of molecules (measured as fea-
tures identified that better discriminate between GPCR
ligands and background compounds), indicating that atom
typing is beneficial in classifications of this type.
Algorithmically, the subgraph miner Gaston was used
that was previously successfully applied to molecular
datasets, details of which can be found in the original publi-
cations [53] and the significance of molecular substructures
was measured as a the likelihood of a feature distribution
occurring by chance, known also as the ‘p-value’ of a distri-
bution (as defined in [54]) and the substructure with the low-
est p-value was considered the most important one. Apart
from the p-value of a substructure, an important parameter in
frequent subgraph mining is the minimum support value,
which is the fraction of molecules in a given dataset that
should possess a particular substructure in order to be con-
sidered for further analysis. Lowering the minimum support
will result in a larger number of substructures and vice versa
– and given the large number (100,000s to millions) of sub-
structures that can be potentially detected in a single mole-
cule it is apparent how important a reasonable choice of this
parameter is. The minimum support value was chosen em-
pirically, resulting in practice in support values between 10%
and 30% for the datasets used in this study.
As important as the molecular representation is also the
choice of a suitable database, and in our study GPCR ligands
were collected from the GLIDA and hGPCR-lig databases.
The set from GLIDA consisted of 22,122 ligands for human,
mouse and rat receptors, while hGPCR-lig contained 17,908
GPCR ligands from literature as well as the MDL Drug Data

8 Current Topics in Medicinal Chemistry, 2011, Vol. 11, No. 16 van der Horst et al.
Report (MDDR) database. These two sets were compared
against a control set, namely 15,993 compounds from
ChemBridge’s DIVERSet screening collection and each da-
tabase was represented by a random set of 5,000 ligands, this
limitation being due to the computational expense of the
methods used here. Targets were arranged into a hierarchy of
subfamilies, families, and classes which is shown in Fig. (7),
and which originates from GPCRDB. For further details the
reader is referred to the original study [48].

















Fig. (7). Schematic drawing of the ligand bioactivity classes sub-
structure mining was applied to (curly brackets indicate the sets that
were compared against each other). ‘Background’ – The Chem-
Bridge DIVERSet, ‘All’ – all GPCR ligands found in GLIDA,
‘Aminergic’ – all aminergic receptor ligands, ‘
\ ’ – Adrenocep-
tors, ‘D’ – Dopamine receptors, ‘H’ – Histamine, ‘M’ – Muscarinic
Acetylcholine receptors, ‘5-HT’ – Serotonin receptors, ‘
’ –
-
adrenoceptors, ‘’ – -adrenoceptors, ‘1-3’ – -adrenoceptor sub-
types 1 to 3. (Reprinted with permission from [48])
The different levels of the target hierarchy are all amena-
ble to substructure mining, and while various studies have
been performed which reproduce known data (such as the
importance of positively charged nitrogens for class A
GPCR ligands relevant for binding to an aspartate residue in
the receptor), also more detailed results have been obtained
from this analysis. Here only one example shall be com-
mented on, namely for dopamine receptor ligands for which
the outcome of the analysis is represented in Fig. (8).
For the dopamine receptor ligands, two types of specific
substructures were identified Fig. (8) that are characteristic
for this type of bioactivity. Here, the first characteristic sub-
structure is present in 30% of the ligands and it consists of a
chain of four or five aromatic atoms, connected to a tertiary
nitrogen atom via a methyl linker. The second substructure,
which is present in 12% of the ligands, consists of two aro-
matic chains which can be five or six atoms in length, and
which are linked via a heteroatom (nitrogen or oxygen) con-
nected to N-methylethyleneamine, while the terminal nitro-
gen of this linker may be substituted by an ethyl group. As
can be seen in Fig. (8), the piperazine ring is part of those
two characteristic substructures, although not in its entirety –
the interpretation is that variations of this ring are possible
when designing dopaminergic drugs, so that this feature does
not remain static throughout the dataset. Similarly, the aro-
matic chains in both substructures are able to overlap with
various types of aromatic systems; hence, it is the aromatic
character of the system that is conserved in this feature, and
not the precise nature of the ring system present. Another
example of novel substructures conferring activity were
fused 5, 6 bicyclic ring systems in serotonergic ligands; these
may steer synthetic chemistry in novel directions when de-
signing future bioactive molecules.
As for future research avenues, currently only the 2D
graph of a molecule is considered for analysis, so any geo-
metric information (such as chirality) is not taken into ac-
count. Also, the p-value is used to sort the substructures ac-
cording to significance; however in practice measures such
as enrichment of a compound bioactivity class might be of
more relevance. Also it is important to also look at substruc-
tures frequent in both data sets when using the information










Fig. (8). Common motif and example substructures for most significant substructures of the dopamine receptor ligands, in aromatic atoms
and bonds representation. An example drug that has motif I is clozapine, an antipsychotic agent used in the treatment of schizophrenia. An-
other example for motif I and also for motif II is compound L-745,870, a selective dopamine D4 receptor antagonist. (Reprinted with permis-
sion from [48])

Chemogenomics for Deorphanization and Phenotype Linkage Current Topics in Medicinal Chemistry, 2011, Vol. 11, No. 16 9
for designing novel drugs – while not conferring selectivity,
those features might still be beneficial for compound affin-
ity. Overall we can conclude from this study that our analy-
sis is complementary to employing ‘privileged substructures’
in ligand design, since it is not restricted to existing scaffold
structures. Apparently, elaborate chemical representations
add substantial value when searching for structural features
typical for active compounds. This enables the user in addi-
tion to detect bioisosteres of chemical groups which can be
employed in the prospective design of ligands with a desired
bioactivity profile.
PROTEOCHEMOMETRICS APPROACHES FOR
EXTRAPOLATING BIOACTIVITY TO RELATED
TARGETS
As mentioned above, the terms ‘chemogenomics’ and
‘proteochemometrics’ are not very different in nature, and
from the experience of the authors the following statement is
reasonable, that in general ‘proteochemometrics models’ are
usually generated based on ‘chemogenomics data’. Proteo-
chemometric models, as opposed to conventional SAR mod-
els, add a target descriptor in addition to the ligand descrip-
tor to the model, in order to use bioactivity information from
related targets to make better predictions where there are
data points known for a target; but also to enable extrapola-
tion of bioactivity data to novel targets – such as mutants of
viral enzymes, or related receptor subtypes (for a recent re-
view see [55] and for related research employing so-called
‘signature descriptors’ see [56]).
Originally this method was intended to improve predic-
tion capabilities on a series of targets where data points were
given [57, 58]. However, it is only a small step expanding
this method to enable the prediction of specificity by inclu-
sion of highly similar targets; in this case the model is then
predicting bioactivities for targets where no data points are
known. This idea is illustrated in Fig. (9) on a hypothetical
dataset that consist of ligand A and very similar ligands B1
and B2, as well as target 1 and very similar targets 2A and
2B. In conventional QSAR models, for every target a sepa-
rate bioactivity model needs to be constructed, and no ex-
trapolation of bioactivities between targets is possible. How-
ever, proteochemometric modeling takes into account that
targets 2A and 2B are similar (e.g. with respect to the shape
and properties of the binding site), hence an approximate
activity also for ligand B2 on target 2A can be predicted, and
also for ligand B1 on target 2B. Both target 1 as well as
ligand A are more dissimilar from the rest of the data, and
here extrapolation is generally possible as a function of
ligand as well as target similarity.
The observation that model extrapolation is possible not
only as a function of ligand similarity, but also as a function
of target similarity, extends the previous ‘applicability do-
main’ concept [59-62]known from QSAR to the biological
domain. We can illustrate it with experimental data from our
group on mutants of viral targets, namely of a set of 14 mu-
tants of HIV reverse transcriptase (manuscript under prepara-
tion). Shown in Fig. (10) is the performance of proteo-
chemometrics modeling in ‘leave-one-sequence-out experi-
ments’, as measured by root mean squared error (RMSE).
The model trained without sequence 8 is seen to perform
inadequately in the validation, most likely due to the under-
estimated impact of the single backbone changing mutation
that is not present in the other mutants. Hence, in proteo-
chemometric modeling the ‘applicability domain’ concept
previously established for the ligand side needs also be ap-
plied to the target side when considering the ability of mod-
els to extrapolate to related protein targets.







Fig. (9). Illustration of proteochemometric modeling on a hypo-
thetical dataset that consist of ligand A and very similar ligands B1
and B2, as well as target 1 and very similar targets 2A and 2B. In
conventional QSAR models, for every target a separate bioactivity
model needs to be constructed, and no extrapolation of bioactivities
between targets is possible. However, proteochemometric modeling
takes into account that targets 2A and 2B are similar (e.g. with re-
spect to the shape and properties of the binding site), hence an ap-
proximate activity for ligand B2 on target 2A can also be predicted,
as well as (with probably more error) an activity of ligand B2 on
target 1. Target 1 as well as ligand A are more dissimilar from the
rest of the data, and here extrapolation is generally possible as a
function of ligand- as well as target similarity.

Furthermore, we applied proteochemometric modeling to
predict the activity of chemical compounds on the four
adenosine receptors by addition of a target description, based
on binding site similarity as this binding site is well known
from a recent crystal structure [63]. We used SCFP4 finger-
prints[64], which were shown to capture more information
than many other methods with respect to their respective
bioactivities[65], together with properties of 32 residues lin-
ing the binding site taken from the AAindex database, to
describe our ligand-target complex. Support Vector Ma-
chines with real-valued output predictions as implemented in
PipelinePilot Student Edition 6.1 were then trained on
known Adenosine Receptor Ligands from ChEMBL. In or-
der to validate the model, we spiked 4,556 random com-
pounds from ZINC with 43 known high-affinity compounds
from our in-house GIFT/LUF compound datasets, and
ranked all ligands using the A2A output of the proteo-
chemometric model predictions to select compounds to
evaluate predictions on this external test set. The ranking of
compounds is shown in Fig. (11), and it can be seen that the
top 14 compounds are true positives (as measured against
any Adenosine receptor subtype), while a total of 37 com-
pounds from our in-house data set were among the 100 high-
est predicted compounds of the entire data set containing
4,599 data points. The three lowest predicted compounds
were found back at position 249, 935 and 1081. The highest
predicted compound was LUF5957 with a predicted pKi of
9.02 and an experimentally determined pKi of 9.14. Given
the satisfying performance of this model, we employed it
also to select novel compounds from supplier databases that

10 Current Topics in Medicinal Chemistry, 2011, Vol. 11, No. 16 van der Horst et al.












Fig. (10). Performance of proteochemometrics modeling in ‘leave-one-sequence-out experiments’, as measured by root mean squared error
(RMSE). The model trained without sequence 8 is seen to perform inadequately in the validation, most likely due to the underestimated im-
pact of the single backbone changing mutation that is not present in the other mutants. Hence, in proteochemometric modeling the ‘applicabil-
ity domain’ concept previously established for the ligand side also needs to be applied to the target side when considering the ability of mod-
els to extrapolate to related protein targets.













Fig. (11). The prediction of a random Zinc decoy set (1139 compounds per receptor) spiked with high affinity compounds active against any
Adenosine receptor (43 in total) from our in-house database, using the A2A bioactivity model. Of the 43 high affinity compounds, 33 were
predicted in the top 50 and 37 in the top 100. The highest predicted decoy compound ranked 14th with a pKi of 7.90.

are currently being evaluated in-house in truly prospective
testing mode.
LINKING PHENOTYPIC SPACE INTO CHEMOGE-
NOMICS ANALYSES
Chemogenomics analyses in their original form are at-
tempting to link ligand chemical space to target bioactivity
space; however, in most cases until now this link has been
restricted to bioactivity against a protein target typically re-
flected in a binding constant, inhibition constant and so
forth. In practice though most often the modulation of a
more complex biological system, such as on an organ or or-
ganism level, is of prime importance and there is no reason
why chemogenomics analyses should not include more com-
plex biological variables than the target similarity by itself.
Recent ambitious analyses though are trying to link ligand
chemical space, via protein target space, to phenotypic space
and some examples of those applications shall be presented
in the following.
One simple observation when dealing with either organ-
isms or cell systems is that they can die. While on an organ-
ism level this can be a disturbing event, also on the cellular
level this can lead to much distress on the side of the
screener when dealing with cell-based assay systems. One
very common assay type in pharmaceutical industry are re-
porter gene assays; and in cases where a decrease in signal
(reporter gene expression) is used as a positive readout,
without normalizing for the number of cells the signal is
taken from cell death can lead to false-positive readouts, and
compounds falsely flagged as ‘hits’ in a reporter gene assay
screen. In our work at Novartis, we employed protein target
prediction models [13,66] for a variety of purposes, and one

Chemogenomics for Deorphanization and Phenotype Linkage Current Topics in Medicinal Chemistry, 2011, Vol. 11, No. 16 11
of them was the analysis of frequent hitters in reporter gene
assays[67], where we assumed a common, underling reason
behind them hitting in a large number of assays in parallel.
Hence, we employed target prediction models on compounds
showing bioactivity in many assays, and it was indeed found
that kinases involved in the cell cycle were very common
targets of frequent hitters [67] – thus, providing a rational
explanation, based on chemogenomics data, for an appar-
ently positive signal in a phenotypic screen.
While cell-based phenotypes are very simple phenotypes,
we can also start to make the crucial step to humans and ad-
verse drug reactions observed upon compound administra-
tion. Recently publications have appeared which go back
from observed adverse drug reactions to a common underly-
ing mechanism of action [68], and in our studies we focused
on a chemogenomics-based target prediction of compounds
with an adverse drug reaction as annotated in the World
Drug Index [52]. For every set of compounds with an ad-
verse drug reaction, targets were predicted, and enriched
targets in a particular compound class were then statistically
associated with a particular adverse reaction; and hence, they
might also be associated with this adverse reaction on a
mechanistic level. An illustration is shown in Fig. (12),
showing the correlation between ligand chemistry active
against proteins and the ligand chemistry causing adverse
drug reactions (phenotypic information). Both target space
and phenotype space have thousands of dimensions, hence a
chemogenomics analysis of the underlying data is potentially
able to unearth novel relationships between mechanistic and
phenotypic space, aiding both drug discovery as well as the
analysis of adverse drug reactions.
A similar analysis was performed from a different type of
data[69], namely high-content screening data, where parts of
the cell are stained, and then geometrical features of the cell
are observed after administration of a compound [70,71].
The essentials of the analysis are shown in Fig. (13), relating
ligand chemistry to the phenotypic response observed, as
well as the protein targets predicted to be hit by the ligand.
An analysis of this type relates a phenotypic observation to
an explanation (mode of action hypothesis), with the chemi-
cal structure being the link between both. It can clearly be
observed that all three types of information (chemistry, phe-
notype, biology) are important to consider when studying
behavior of a compound, since none of them is correlated
well enough to any other.
More recently, also the concept of applying multiple in-
terventions in parallel to biological systems in a coordinated
manner has found considerable interest, both in the pharma-
ceutical area [72-74] as well as in more fundamental research
[75,76]. This is based partly on our increasing understanding
of single interventions, now making partially also possible to
engineer desired combinations of interventions, but even
more so this is due to the realization that multiple targets are
needed to modulate biological networks in the desired man-
ner due to effects such as redundant mechanisms [77,78],
that render a cell more stable in more adverse conditions.
From the chemical side, the concept of ‘chemical genetics’ is
of much relevance here [5], where the biological concepts of
genetics (such as gene knock-out experiments) are mimicked
by compound application. This is not meant to resemble ge-
netics completely, and in fact chemical genetics experiments
differ much from their genetic siblings, since they allow in-
terventions in a dose- and time-dependent manner. Also,
protein surfaces even of inhibited proteins are still available
to mediate protein-protein contacts, which is not the case if a
protein is not expressed at all in the first place.
From the experimental side, the discovery of ligands with
the desired bioactivity profile requires novel techniques such
as diversity-oriented synthesis (DOS) to firstly explore
chemistry potentially active against the desired targets
[79,80], followed by more conventional optimization to-
wards the required bioactivity profile. However, as known
from drug discovery, serendipity is likely to play a promi-
nent role in the area for the foreseeable future [81].
CONCLUSIONS
We are currently witnessing an ever increasing amount of
publicly accessible bioactivity data, and this is coinciding
with concepts such as polypharmacology which are currently
being recognized for their importance in pharmaceutical re-
search in academia as well as industry. The question of how
to design bioactive matter against a single target is non-
trivial; it is easy to imagine that designing chemistry with the







Fig. (12). Correlation between ligand chemistry active against proteins and the ligand chemistry causing adverse drug reactions (phenotypic
information; darker colours indicate higher correlation). Both target space and phenotype space have thousands of dimensions, hence a che-
mogenomics analysis of the underlying data is potentially able to unearth novel relationships between mechanistic and phenotypic space,
aiding both drug discovery as well as the analysis of adverse drug reactions.

12 Current Topics in Medicinal Chemistry, 2011, Vol. 11, No. 16 van der Horst et al.
right bioactivity profile against multiple targets is even more
demanding. Hence, it becomes clear that we need the che-
mogenomics data at our disposal today to make wiser deci-
sions regarding the design of bioactive matter in the future,
by employing algorithms that have partly been developed
already, but which for the most part still have to be con-
ceived in the future. We have tens of millions of bioactivity
data points available today – now we have to develop ways
to make proper use of them.
ACKNOWLEDGMENTS
EvdH, APIJ and AB thank Dutch Top Institute Pharma
(TI Pharma) for support of project number D1-105 during
the receptor deorphanization and chemical representations
parts of this work. AB thanks the Education Office of the
Novartis Institutes for BioMedical Research for supporting
the chemical representation and phenotype work of this
manuscript and current funding. GJPvW thanks Tibotec for
funding. DRS and WRJDG thank EPSRC, BBSRC, MRC
and the Wellcome and Newman Trusts for funding.
REFERENCES
[1] Bredel M.; Jacoby, E. Chemogenomics: an emerging strategy for
rapid target and drug discovery. Nat. Rev. Genet., 2004, 5, 262-275.
[2] Jacoby E. A novel chemogenomics knowledge-based ligand design
strategy - Application to G protein-coupled receptors. Quant.
Struct.-Act. Relat., 2001, 20, 115-123.
[3] Caron P. R.; Mullican M. D.; Mashal R. D.; Wilson K. P.; Su M. S.
Murcko M. A. Chemogenomic approaches to drug discovery. Curr.
Opin. Chem. Biol., 2001, 5, 464-470.
[4] Schuffenhauer A.; Floersheim P.; Acklin P. Jacoby E. Similarity
metrics for ligands reflecting the similarity of the target proteins. J.
Chem. Inf. Comput. Sci., 2003, 43, 391-405.
[5] Spring D. R. Chemical genetics to chemical genomics: small mole-
cules offer big insights. Chem. Soc. Rev., 2005, 34, 472-482.
[6] Stockwell B. R. Chemical genetics: Ligand-based discovery of
gene function. Nat. Rev. Genet., 2000, 1,116-125.
[7] Wishart D. S.; Knox C.; Guo A. C.; Shrivastava S.; Hassanali M.;
Stothard P.; Chang Z. Woolsey J. DrugBank: a comprehensive re-
source for in silico drug discovery and exploration. Nucl. Acids
Res., 2006, 34, D668-672.
[8] Liu T.; Lin Y.; Wen X.; Jorissen R. N.; Gilson M. K. BindingDB: a
web-accessible database of experimentally determined protein-
ligand binding affinities. Nucl. Acids Res., 2007, 35, D198-201.
[9] Roth B. L.; Lopez E.; Beischel S.; Westkaemper R. B. Evans J. M.
Screening the receptorome to discover the molecular targets for
plant-derived psychoactive compounds: a novel approach for CNS
drug discovery. Pharmacol. Ther., 2004, 102, 99-110.
[10] PubChem - Part of the NIH Molecular Libraries Roadmap Initia-
tive. http://pubchem.ncbi.nlm.nih.gov/.
[11] http://chembank.broad.harvard.edu/. C. ChemBank.
http://chembank.broad.harvard.edu/.
[12] Williams A. J. Public chemical compound databases. Curr. Opin.
Drug Discov. Devel., 2008, 11, 393-404.
[13] Jenkins J. L.; Bender A., Davies J. W. In silico target fishing: Pre-
dicting biological targets from chemical structure. Drug Discov.
Today Technol., 2007, 3, 413-421.
[14] ChEMBL database, accessible at http://www.ebi.ac.uk/chembl/.
[15] Bender A. Databases: Compound bioactivities go public. Nat.
Chem. Biol., 2010, 6, 309-309.
[16] Kuhn M.; Campillos M.; Letunic I.; Jensen L. J. Bork P. A side
effect resource to capture phenotypic effects of drugs. Mol. Syst.
Biol., 2010, 6, 343.
[17] Mestres J.; Gregori-Puigjane E.; Valverde S.; Sole R. V. The to-
pology of drug-target interaction networks: implicit dependence on
drug properties and target families. Mol. Biosyst., 2009, 5, 1051-
1057.
[18] Mestres J.; Gregori-Puigjane E.; Valverde S.; Sole R. V. Data
completeness--the Achilles heel of drug-target networks. Nat. Bio-
technol., 2008, 26, 983-984.
[19] Csermely P.; Agoston V. Pongor S. The efficiency of multi-target
drugs: the network approach might help drug design. Trends Phar-
macol. Sci., 2005, 26, 178-182.
[20] Morphy R.; Rankovic Z. Designed multiple ligands. An emerging
drug discovery paradigm. J. Med. Chem., 2005, 48, 6523-6543.
[21] Bender A.; Jenkins J. L.; Glick M.; Deng Z.; Nettles J. H., Davies
J. W. “Bayes Affinity Fingerprints” improve retrieval rates in vir-
tual screening and define orthogonal bioactivity space: When are
multitarget drugs a feasible concept? J. Chem. Inf. Model., 2006,
46, 2445-2456.
[22] Paolini G. V.; Shapland R. H.; van Hoorn W. P.; Mason J. S.; Hop-
kins A. L. Global mapping of pharmacological space. Nat. Bio-
technol., 2006, 24, 805-815.











Fig. (13). Analysis of ligand chemical space, in relation to the phenotypic response observe in a microscopy-based high-content screen, as
well as the protein targets predicted to be hit by the ligand (ligand numbers taken from the original publication[69]). An analysis of this type
relates a phenotypic observation to an explanation (mode of action hypothesis), with the chemical structure being the link between both. It
can clearly be observed that all three types of information (chemistry, phenotype, biology) are important to consider when studying the be-
haviour of a compound, since none of them is correlated well enough to any other. (Reprinted with permission from [69].)

Chemogenomics for Deorphanization and Phenotype Linkage Current Topics in Medicinal Chemistry, 2011, Vol. 11, No. 16 13
[23] Fliri A. F.; Loging W. T.; Thadeio P. F.; Volkmann R. A. Analysis
of drug-induced effect patterns to link structure and side effects of
medicines. Nat. Chem. Biol., 2005, 1, 389-397.
[24] Keiser M. J.; Roth B. L.; Armbruster B. N.; Ernsberger P.; Irwin J.
J. Shoichet B. K. Relating protein pharmacology by ligand chemis-
try. Nat. Biotechnol., 2007, 25, 197-206.
[25] Gregori-Puigjane E. Mestres J. A ligand-based approach to mining
the chemogenomic space of drugs. Comb Chem High Throughput
Screen, 2008, 11, 669-676.
[26] Kauvar L. M.; Higgins D. L.; Villar H. O.; Sportsman J. R.;
Engqvist-Goldstein A.; Bukar R.; Bauer K. E.; Dilley H.; Rocke D.
M. Predicting ligand binding to proteins by affinity fingerprinting.
Chem. Biol., 1995, 2, 107-118.
[27] Briem H.; Lessel U. In vitro and in silico affinity fingerprints:
Finding similarities beyond structural classes. Perspect. Drug Dis-
cov. Des., 2000, 20, 231-244.
[28] Strombergsson H.; Daniluk P.; Kryshtafovych A.; Fidelis K.; Wik-
berg J. E.; Kleywegt G. J. Hvidsten T. R. Interaction model based
on local protein substructures generalizes to the entire structural
enzyme-ligand space. J. Chem. Inf. Model., 2008, 48, 2278-2288.
[29] Roche O.; Schneider P.; Zuegge J.; Guba W.; Kansy M.; Alanine
A.; Bleicher K.; Danel F.; Gutknecht E. M.; Rogers-Evans M.;
Neidhart W.; Stalder H.; Dillon M.; Sjogren E.; Fotouhi N.; Gilles-
pie P.; Goodnow R.; Harris W.; Jones P.; Taniguchi M.; Tsujii S.;
von der Saal W.; Zimmermann G. Schneider G. Development of a
virtual screening method for identification of "frequent hitters" in
compound libraries. J. Med. Chem., 2002, 45, 137-142.
[30] Mestres J. Computational chemogenomics approaches to system-
atic knowledge-based drug discovery. Curr. Opin. Drug Discov.
Devel., 2004, 7, 304-313.
[31] Rognan D. Chemogenomic approaches to rational drug design. Br.
J. Pharmacol., 2007, 152, 38-52.
[32] Klabunde T. Chemogenomic approaches to drug discovery: similar
receptors bind similar ligands. Br. J. Pharmacol., 2007, 152, 5-7.
[33] Marechal E. Chemogenomics: a discipline at the crossroad of high
throughput technologies, biomarker research, combinatorial chem-
istry, genomics, cheminformatics, bioinformatics and artificial in-
telligence. Comb. Chem. High Throughput. Screen, 2008, 11, 582.
[34] Jacoby E. Mozzarelli A. Chemogenomic strategies to expand the
bioactive chemical space. Curr. Med. Chem., 2009, 16, 4374-4381.
[35] Bender A.; Young D. W.; Jenkins J. L.; Serrano M.; Mikhailov D.;
Clemons P. A.; Davies J. W. Chemogenomic data analysis: predic-
tion of small-molecule targets and the advent of biological finger-
print. Comb. Chem. High-Throughput Screen, 2007, 10, 719-731.
[36] Doddareddy M. R.; van Westen G. J. P.; van der Horst E.; Peiron-
cely J. E.; Corthals F.; Ijzerman A. P.; Emmerich M.; Jenkins J. L.;
Bender A. Chemogenomics: Looking at Biology through the Lens
of Chemistry. Stat. Anal. Data Mining, 2009, 2, 149-160.
[37] Vieth M.; Higgs R. E.; Robertson D. H.; Shapiro M.; Gragg E. A.
Hemmerle H. Kinomics-structural biology and chemogenomics of
kinase inhibitors and targets. Biochim. Biophys. Acta, 2004, 1697,
243-257.
[38] Karaman M. W.; Herrgard S.; Treiber D. K.; Gallant P.; Atteridge
C. E.; Campbell B. T.; Chan K. W.; Ciceri P.; Davis M. I.; Edeen
P. T.; Faraoni R.; Floyd M.; Hunt J. P.; Lockhart D. J.; Milanov Z.
V.; Morrison M. J.; Pallares G.; Patel H. K.; Pritchard S.; Wodicka
L. M.; Zarrinkar P. P. A quantitative analysis of kinase inhibitor se-
lectivity. Nat. Biotechnol., 2008, 26, 127-132.
[39] Bantscheff M.; Eberhard D.; Abraham Y.; Bastuck S.; Boesche M.;
Hobson S.; Mathieson T.; Perrin J.; Raida M.; Rau C.; Reader V.;
Sweetman G.; Bauer A.; Bouwmeester T.; Hopf C.; Kruse U.;
Neubauer G.; Ramsden N.; Rick J.; Kuster B.; Drewes G. Quantita-
tive chemical proteomics reveals mechanisms of action of clinical
ABL kinase inhibitors. Nat. Biotechnol., 2007, 25, 1035-1044.
[40] Bamborough P.; Drewry D.; Harper G.; Smith G. K.; Schneider K.
Assessment of chemical coverage of kinome space and its implica-
tions for kinase drug discovery. J. Med. Chem., 2008, 51, 7898-
7914.
[41] Bernasconi P.; Chen M.; Galasinski S.; Popa-Burke I.; Bobasheva
A.; Coudurier L.; Birkos S.; Hallam R.; Janzen W. P. A Chemoge-
nomic analysis of the human proteome: Application to enzyme
families. J. Biomol. Screen., 2007, 12, 972-982.
[42] Jacob L.; Hoffmann B.; Stoven V.; Vert J. P. Virtual screening of
GPCRs: an in silico chemogenomics approach. BMC Bioinformat-
ics, 2008, 9, 363.
[43] Bock J. R.; Gough D. A. Virtual screen for ligands of orphan g
protein-coupled receptors. J. Chem. Inf. Model., 2005, 45,402-
1414.
[44] Weill N.; Rognan D. Development and validation of a novel pro-
tein-ligand fingerprint to mine chemogenomic space: application to
G protein-coupled receptors and their ligands. J. Chem. Inf. Model.,
2009, 49, 1049-1062.
[45] Van der Horst E.; Peironcely J. E.; Ijzerman A. P.; Beukers M. W.;
Lane J. R.; van Vlijmen H. W. T.; Emmerich M. T. M.; Okuno Y.;
Bender A. A novel chemogenomics analysis of G protein-coupled
receptors (GPCRs) and their ligands: a potential strategy for recep-
tor de-orphanization. BMC Bioinformatics, 2010, 11, 316.
[46] Gloriam D. E.; Foord S. M.; Blaney F. E.; Garland S. L. Definition
of the G protein-coupled receptor transmembrane bundle binding
pocket and calculation of receptor similarities for drug design. J.
Med. Chem., 2009, 52, 4429-4442.
[47] Surgand J. S.; Rodrigo J.; Kellenberger E.; Rognan D. A chemoge-
nomic analysis of the transmembrane binding cavity of human G-
protein-coupled receptors. Proteins, 2006, 62, 509-538.
[48] van der Horst E.; Okuno Y.; Bender A.; AP I. J. Substructure min-
ing of GPCR ligands reveals activity-class specific functional
groups in an unbiased manner. J. Chem. Inf. Model., 2009, 49, 348-
360.
[49] Bender A.; Jenkins J. L.; Scheiber J.; Sukuru S. C.; Glick M.; Da-
vies J. W. How similar are similarity searching methods? A princi-
pal component analysis of molecular descriptor space. J. Chem. Inf.
Model., 2009, 49, 108-119.
[50] Bender A.; Glen R. C. A Discussion of measures of enrichment in
virtual screening: comparing the information content of descriptors
with increasing levels of sophistication. J. Chem. Inf. Model., 2005,
45, 1369-1375.
[51] Bender A.; Glen R. C. Molecular similarity: a key technique in
molecular informatics. Org. Biomol. Chem., 2004, 2, 3204-3218.
[52] Bender A.; Scheiber J.; Glick M.; Davies J. W.; Azzaoui K.;
Hamon J.; Urban L.; Whitebread S. Jenkins J. L. Analysis of phar-
macology data and the prediction of adverse drug reactions and off-
target effects from chemical structure ChemMedChem, 2007, 2,
861-873.
[53] Kazius J.; Nijssen S.; Kok J.; Back T. Ijzerman A. P. Substructure
mining using elaborate chemical representation. J. Chem. Inf.
Model., 2006, 46, 597-605.
[54] Kazius J.; McGuire R. Bursi R. Derivation and validation of toxi-
cophores for mutagenicity prediction. J. Med. Chem., 2005, 48,
312-320.
[55] van Westen G. J. P.; Wegener J. K.; Ijzerman A. P.; Van Vlijmen
H. W. T.; Bender A. Proteochemometric modeling as a tool for de-
signing selective compounds and extrapolating to novel targets.
Med. Chem. Comm., 2011, 2, 16-30
[56] Faulon J. L.; Misra M.; Martin S.; Sale K., Sapra R. Genome scale
enzyme-metabolite and drug-target interaction predictions using the
signature molecular descriptor. Bioinformatics, 2008, 24, 225-233.
[57] Lapinsh M.; Prusis P.; Gutcaits A.; Lundstedt T., Wikberg J. E. S.
Development of proteo-chemometrics: a novel technology for the
analysis of drug-receptor interactions. Biochim. Biophys. Acta-Gen.
Subj., 2001, 1525, 180-190.
[58] Lapinsh M.; Prusis P.; Lundstedt T. Wikberg J. E. S. Proteo-
chemometrics modeling of the interaction of amine G-protein cou-
pled receptors with a diverse set of ligands. Mol. Pharmacol., 2002,
61, 1465-1475.
[59] Dragos H.; Gilles M. Alexandre V. Predicting the predictability: a
unified approach to the applicability domain problem of QSAR
models. J. Chem. Inf. Model., 2009, 49, 1762-1776.
[60] Tetko I. V.; Sushko I.; Pandey A. K.; Zhu H.; Tropsha A.; Papa E.;
Oberg T.; Todeschini R.; Fourches D.; Varnek A. Critical assess-
ment of QSAR models of environmental toxicity against Tetrahy-
mena pyriformis: focusing on applicability domain and overfitting
by variable selection. J. Chem. Inf. Model., 2008, 48, 1733-1746.
[61] Weaver S.; Gleeson M. P. The importance of the domain of appli-
cability in QSAR modeling. J Mol. Graph. Model., 2008, 26, 1315-
1326.
[62] Schroeter T.; Schwaighofer A.; Mika S.; Laak A. T.; Suelzle D.;
Ganzer U.; Heinrich N.; Muller K. R. Machine learning models for
lipophilicity and their domain of applicability. Mol. Pharm., 2007,
4, 524-538.
[63] Jaakola V. P.; Griffith M. T.; Hanson M. A.; Cherezov V.; Chien E.
Y.; Lane J. R.; Ijzerman A. P.; Stevens R. C. The 2.6 angstrom

14 Current Topics in Medicinal Chemistry, 2011, Vol. 11, No. 16 van der Horst et al.
crystal structure of a human A2A adenosine receptor bound to an
antagonist. Science, 2008, 322, 1211-1217.
[64] Rogers D., Hahn M. Extended-connectivity fingerprints. J. Chem.
Inf. Model., 2010, 50, 742-754.
[65] Glen R. C.; Bender A.; Arnby C. H.; Carlsson L.; Boyer S.; Smith
J. Circular fingerprints: Flexible molecular descriptors with appli-
cations from physical chemistry to ADME. IDrugs, 2006, 9, 199-
204.
[66] Nettles J. H.; Jenkins J. L.; Bender A.; Deng Z.; Davies J. W.;
Glick M. Bridging chemical and biological space: "target fishing"
using 2D and 3D molecular descriptors. J. Med. Chem., 2006, 49,
6802-6810.
[67] Crisman T. J.; Parker C. N.; Jenkins J. L.; Scheiber J.; Thoma M.;
Kang Z. B.; Kim R.; Bender A.; Nettles J. H.; Davies J. W.; Glick
M. Understanding False Positives in Reporter Gene Assays: in
silico Chemogenomics Approaches to Prioritize Cell-based HTS
Data. J. Chem Inf. Model., 2007, 47, 1319-1327.
[68] Campillos M.; Kuhn M.; Gavin A. C.; Jensen L. J.; Bork P. Drug
target identification using side-effect similarity. Science, 2008, 321,
263-266.
[69] Young D. W.; Bender A.; Hoyt J.; McWhinnie E.; Chirn G. W.;
Tao C. Y.; Tallarico J. A.; Labow M.; Jenkins J. L.; Mitchison T.
J.; Feng Y. Integrating high-content screening and ligand-target
prediction to identify mechanism of action. Nat. Chem. Biol., 2008,
4, 59-68.
[70] Feng Y.; Mitchison T. J.; Bender A.; Young D. W. Tallarico J. A.
Multi-parameter phenotypic profiling: using cellular effects to
characterize small-molecule compounds. Nat. Rev. Drug Discov.,
2009, 8, 567-578.
[71] Kummel A.; Gabriel D.; Parker C. N.; Bender A. Computational
methods to support high-content screening: from compound selec-
tion and data analysis to postulating target hypotheses. Exp. Op.
Drug Disc., 2009, 4, 5-13.
[72] Jia J.; Zhu F.; Ma X.; Cao Z.; Li Y.; Chen Y. Z. Mechanisms of
drug combinations: interaction and network perspectives. Nat. Rev.
Drug Discov., 2009, 8, 111-128.
[73] Lehar J.; Krueger A. S.; Zimmermann G. R.; Borisy A. A. Thera-
peutic selectivity and the multi-node drug target. Discov. Med.,
2009, 8,185-190.
[74] Lehar J.; Krueger A. S.; Avery W.; Heilbut A. M.; Johansen L. M.;
Price E. R.; Rickles R. J.; Short G. F3rd.; Staunton J. E.; Jin X.; Lee
M. S.; Zimmermann G. R.; Borisy A. A. Synergistic drug combina-
tions tend to improve therapeutically relevant selectivity. Nat. Bio-
technol., 2009, 27, 659-666.
[75] Lehar J.; Stockwell B. R.; Giaever G.; Nislow C. Combination
chemical genetics. Nat. Chem. Biol., 2008, 4, 674-681.
[76] Nelander S.; Wang W.; Nilsson B.; She Q. B.; Pratilas C.; Rosen
N.; Gennemark P.; Sander C. Models from experiments: combina-
torial drug perturbations of cancer cells. Mol. Syst. Biol., 2008, 4,
216.
[77] Kitano H. A robustness-based approach to systems-oriented drug
design. Nat. Rev. Drug Discov., 2007, 6, 202-210.
[78] Kartal O.; Ebenhoh O. Ground state robustness as an evolutionary
design principle in signaling networks. PLoS One, 2009,4,e8001.
[79] Spandl R. J.; Bender A.; Spring D. R. Diversity-oriented synthesis;
a spectrum of approaches and results. Org. Biomol. Chem., 2008, 6,
1149-1158.
[80] Fergus S.; Bender A.; Spring D. R. Assessment of structural diver-
sity in combinatorial synthesis. Curr. Op. Chem. Biol, 2005, 9,
304-309.
[81] Kubinyi H. Chance favors the prepared mind--from serendipity to
rational drug design. J. Recept. Signal. Transduct. Res., 1999, 19,
15-39.



Received: June 18, 2010 Revised: August 22, 2010 Accepted: September 02, 2010