How to design multi-target drugs: Target search options in cellular networks

Abstract

Despite improved rational drug design and a remarkable progress in genomic, proteomic and high-throughput screening methods, the number of novel, single-target drugs has fallen far behind expectations during the past decade. Multi-target drugs multiply the number of pharmacologically relevant target molecules by introducing a set of indirect, network-dependent effects. Parallel with this, the low-affinity binding of multi-target drugs eases the constraints of druggability and significantly increases the size of the druggable proteome. These effects tremendously expand the number of potential drug targets and introduce novel classes of multi-target drugs with smaller side effects and toxicity. Here, the authors review the recent progress in this field, compare possible network attack strategies and propose several methods to find target-sets for multi-target drugs.

Full text

Review
10.1517/17460441.2.6.1 © 2007 Informa UK Ltd ISSN 1746-0441 1
1. Introduction: emergence and
rationale of the multi-drug
concept
2. Examples for multi-target
strategies
3. Cellular networks: drug
target maps
4. Multi-target drugs are often
low-affi nity binders
5. Identifi cation of drug targets
using the network approach:
attack strategies
6. Network diseases
7. Target-sets of multi-target
drugs: the help of networks
8. Conclusion
9. Expert opinion: network-based,
smart multi-target drugs
of the future
How to design multi-target
drugs : target search options
in cellular networks
Tamás Korcsmáros , Máté S Szalay , Csaba Böde , István A Kovács &
Péter Csermely †
†Semmelweis University, Department of Medical Chemistry, PO Box 260, H-1444 Budapest 8,
Hungary and Predinet Ltd., Dongo Street 8, H-1149 Budapest, Hungary
Despite improved rational drug design and a remarkable progress in genomic,
proteomic and high-throughput screening methods, the number of novel,
single-target drugs has fallen much behind expectations during the past
decade. Multi-target drugs multiply the number of pharmacologically
relevant target molecules by introducing a set of indirect, network-dependent
effects. Parallel with this, the low-affinity binding of multi-target drugs eases
the constraints of druggability and significantly increases the size of the
druggable proteome. These effects tremendously expand the number of
potential drug targets and introduce novel classes of multi-target drugs with
smaller side effects and toxicity. Here, the authors review the recent progress
in this field, compare possible network attack strategies and propose several
methods to find target-sets for multi-target drugs.
Keywords: antibiotics , disordered proteins , drug targets , fungicides , genomics ,
multi-target drugs , networks , network damage , pesticides , phosphorylation ,
protein kinases , proteomics , signaling networks
Expert Opin. Drug Discov. (2007) 2(6):1-10
1. Introduction: emergence and rationale
of the multi-drug concept
Recent drug development strategies were based on the emergence of potential targets
in genomic and proteomic studies. Therefore, the drug development paradigm
followed at present can be summarized as to: i) find a target of clinical relevance;
ii) identify the ‘best-binder druggable molecule’ by high-throughput screening
(HTS) of large combinatorial libraries and/or by rational drug design based on the
three-dimensional structure of the target; iii) provide a set of proof-of-principle
experiments; and iv) develop a technology platform leading to clinical applications.
However, despite all the considerable drug development efforts undertaken, the
number of successful drugs and novel targets fell significantly behind the expectations
during past decades [1-3] .
A number of novel strategies have been developed to overcome the target shortage
and to add novel classes of drugs to development pipelines. Many of these drug
development directions aim to influence multiple targets in a parallel fashion. One
of the most widespread multiple target approaches, combination therapy, is increas-
ingly used to treat many types of diseases, such as AIDS, atherosclerosis, cancer
and depression. As one of the newly developed combination therapies, ‘multi-target
lead discovery’ is a promising tool for the identification of unexpectedly novel effects
of drug combinations [4-8] . Recently, initial steps have been taken to develop aptamer
combinations against complex sets of targets [9] .
Multiple target strategies have only recently been rediscovered by drug developers.
Snake and spider venoms are both multi-component systems and plants have

How to design multi-target drugs: target search options in cellular networks
2 Expert Opin. Drug Discov. (2007) 2(6)
also developed a combinative strategy to defend themselves
against pathogens. In addition, traditional medicaments and
remedies often contain multi-component extracts of natural
products [10,11] . All these examples show that multiple target
strategies have benefits, which were used as medicaments by
human ancestors several thousand years ago and have
withstood a million-year evolutionary selection.
Agents aiming at only a single target (‘single-hits’) might
not always affect complex systems in the desired way, even
if they completely change the behavior of their immediate
target. Single targets might have ‘back-up’ systems that are
sometimes different enough so as to not respond to the
same drug. Moreover, many cellular networks are robust and
prevent major changes in their outputs, despite dramatic
changes in their constituents [12-15] . These considerations are
independent of whether or not the pharmacologic agent
inhibits or activates its target.
2. Examples for multi-target strategies
Several efficient drugs, such as salicylate, non-steroidal anti-
inflammatory drugs (NSAIDs), metformin, antidepressants,
antineurodegenerative agents and multi-target kinase
inhibitors (such as Imatinib, or the inhibitors of the kinase-
maturating molecular chaperone, Hsp90), affect many targets
simultaneously [7,16-24] . Multi-target antibodies (in forms of
diabodies, triabodies, tetrabodies and recombinant polyclonal
antibodies) are increasingly used in cancer therapy to delay
the development of resistance [25,26] .
It may be found that a large number of terms used at
present to describe ligands have multiple activities: the
words balanced, binary, bivalent, dimeric, dual, mixed or
triple are all used in combination with various suffixes, such as
agonist, antagonist, blocker, conjugate, inhibitor or ligand.
Various pharmacophores may have an increasing overlap,
which gives an almost continuous spectrum starting from
the conjugates (or cleavable conjugates, which are actually
a novel chemical form of combinational therapies) to the
overlapping pharmacophores, until the highly integrated
multi-target drugs ( Figure 1 [7,19] ).
C. Integrated
multi-target
drug
B. Overlapping
pharmacophores
A. Conjugate
Figure 1. Variants of multi-target drugs. A. The increasing
overlap of pharmacophores gives an almost continuous
spect rum start ing from conjugates. B. Via slightly overlapping
pharmacophores. C. Until highly integrated multi-target drugs.
Multi-target drugs offer a magnification of the ‘sweet spot’
of drug discovery [1] , meaning the overlap between pathways,
which are interesting from the pharmacologic point of view,
and the hits of chemical proteomics, which represent those
proteins that can interact with druggable molecules (meaning
small, hydrophobic molecules with a good bioavailability).
The sweet spot represents those few hundred proteins, which are
both parts of these interesting pathways and are druggable [1] .
The option to allow indirect effects via network-contacts
of multi-target drugs expands the first circle, as the number of
those proteins, which are indirectly related to existing targets
of pharmacologically important pathways, is by magnitudes
greater than the number of the targets themselves. On the
other hand, the low-affinity binding of multi-target drugs
enlarges the second circle, as it eases the constraints of
druggability. Small, hydrophobic molecules bind to only a
small subset of proteins with high affinity. However, the very
same molecules interact with 10 or even 100 times more proteins
with increasingly lower and lower affinity. Here, low affinity
binding describes interactions with dissociation constants in
the higher macromolar or even close to the millimolar range.
Low affinity binding also implies a more transient interaction
(where the off-rate is comparable or higher than the on-rate).
As a result of these combined effects, the sweet spot of drug
discovery may easily become a wide candy-field ( Figure 2 ).
3. Cellular networks: drug target maps
Cellular networks help us to understand the complexity of the
cell. In the network concept, the complex system is perceived
as a set of interacting elements that are bound together by links.
Links usually have a weight, which characterizes their strength
(affinity, or propensity). Links may also be directed links,
when one of the elements has a larger influence than the other
and vice versa. In cellular networks, the interacting molecules
are considered as the elements and their interactions form the
weighted, but not necessarily directed, links of the respective
structural network. Alternatively, directed links may also be
seen as representations of signaling or metabolic processes of
the functional networks in the cell ( Table 1 [27-29] ). Cellular
networks often form small worlds where two elements of the
network are separated by only a few other elements. Networks of
our cells contain hubs, that is, elements which have a large
number of neighbors. These networks can be dissected to over-
lapping modules that form hierarchical communities [30-32] .
However, this summary of the main features of cellular networks
is largely a generalization and needs to be validated through a
critical scrutiny of the data sets, sampling procedures and
methods of data analysis at each network examined [33,34] .
Cellular networks offer a lot of possibilities to point out
their key elements as potential drug targets. An example of
these possibilities, signaling networks have interdigitated path-
ways and multiple layers of cross-talk [35] . Special signaling
elements, such as the PI3 kinase, the Akt kinase or the insulin
receptor substrate family have been called ‘critical nodes’.

Korcsmáros, Szalay, Böde, Kovács & Csermely
Expert Opin. Drug Discov. (2007) 2(6) 3
These ‘critical nodes’ have multiple isoforms and are important
junctions of signaling pathways [36] . Both the bridge elements
of signaling networks, providing cross-talk and the critical
nodes, can be important targets of network-based drug
development. Domain-specific target analysis of protein–
protein interaction networks extends the map of physical
interactions towards functional understanding. Domain-
specific targets offer a larger flexibility and may actually reflect
a family of multiple targets due to the frequent re-use of
domain variants as a result of modular evolution [37,38] .
Elements of metabolic and cytoskeletal networks have also
been analyzed as drug targets [39,40] .
However, present databases of most cellular networks suffer
a lot of uncertainties. Protein–protein interaction networks
have a large number of false-positive entries, which makes the
inclusion and assessment of low-affinity interactions especially
difficult. Moreover, present databases mostly give an averaged
probability of the particular interaction. This does not take
into account if the two proteins are expressed at the same time
or if they are located in the same cellular compartment at the
same time. Most databases do not contain the information of
the ratio of two interacting proteins in the given status of the
particular healthy or the particular sick cell. Literature-derived,
evidence-based databases suffer from the nomenclature and
interpretation problems of the original data. However, recent
advances connect protein–protein interaction databases with
protein structure data, which make both the validation and
prediction of protein–protein interactions more robust
[41-43] .
The above-mentioned problem may be overcome by using
curated databases, which contain only the most valid and most
accurate information. However, these databases will miss
most low-affinity interactions and ∼ 80% of the available
information is lost due to increased scrutiny. As an alternative
approach, the authors keep all information – taking into
account that the author’s database becomes ‘fuzzy’ due to the
inclusion of potentially false data. In this approach to correct
the errors, highly integrated methods for network analysis are
needed, which are able to build in all the above information
and take into account those which are in contradiction with
most of the others. It is advantageous to use these integrated
analytical tools in a ‘zoom-in’ fashion, where the user may
define the integration level of their choice. Low resolution
network maps can be calculated faster and by directing the
user’s mind, showing the most important take-home messages
of the analysis. Zooming in to a high-resolution analysis with
the same, flexible method, will show the refined details of all
available information and give the user a large number of cor-
rect (and false) ideas to think about and to test in experiments
either in the primary ‘hit areas’ of the low-resolution analysis
or the spots of specific interests based on other assumptions.
4. Multi-target drugs are often
low-affi nity binders
The development of a multi-target drug is likely to produce
a drug that interacts with its target (having a lower affinity
than a single-target drug) because it is unlikely that a small,
drug-like molecule can bind to a number of different targets
with equally high affinity. However, low-affinity drug binding
is, apparently, not a disadvantage. For example, memantine
(a drug used to treat Alzheimer’s disease) and other multi-
target non-competitive NMDA receptor antagonists show that
low-affinity, multi-target drugs might have a lower prevalence
and a reduced range of side effects than high-affinity, single-target
A. The current ‘sweet spot’
of drug discovery
B. Target multiplication,
new drug classes
Novel drug
target-sets
Druggable
proteins
Potential
drug targets
Multi-target
drugs
Potential
drug targets
The total
cellular proteome
The total
cellular proteome
Druggable
proteins
Figure 2. From the ‘sweet spot’ of drug discovery to a potential candy-fi eld. Multi-target drugs may magnify the ‘sweet spot’ of
drug discovery [1] to a whole candy-fi eld. A. The overlap between pathways, which is interesting from the pharmacological point of view
and the hits of chemical proteomics, representing those proteins that can interact with druggable molecules, constitutes the ‘sweet spot’
of drug discovery. B. Indirect effects of multi-target drugs expand the number of pharmacologically relevant targets, whereas low-affi nity
binding enlarges the number of druggable molecules.

How to design multi-target drugs: target search options in cellular networks
4 Expert Opin. Drug Discov. (2007) 2(6)
drugs [20,44] . The recent suggestion to use unstructured proteins
as a novel and un-explored field of drug-targets uses the
beneficial effects of low-affinity, but rather specific binding,
which has been shown to be extremely useful in regulation
and signaling [45] . Here again, low affinity binding denotes
interactions with dissociation constants in the higher micro-
molar or even close to the millimolar range. Low affinity
binding also implies a transient interaction, where the off-rate
is comparable to or higher than the on-rate.
Does low-affinity binding predict a low-efficiency? Not
necessarily. Most (> 80%) of the cellular protein, signaling and
transcriptional networks are in a low-affinity or transient ‘weak
linkage’ with each other. In metabolic networks, weak links are
those reactions that have a low flux [29,46] . In this review, the
authors use the term ‘weak linker’ to denote small molecules
and drugs that interact with cellular proteins having a low-
affinity. Thus, most multi-target drugs are weak linkers.
Because most links in cellular networks are weak, a low-affinity
multi-target drug might be sufficient to achieve a significant
modification. A recent paper of B N Ames [47] on the potential
impact of micronutrients on disease development is a good
example of the profound effects emerging from seemingly
minor interactions. Low affinity, imperfect binding allows the
development of special, cooperative binding behavior, which
may lead to a switch-type activation setting a threshold for
various cellular events, such as DNA replication [48,49] .
5. Identifi cation of drug targets using the
network approach: attack strategies
Drug design strategies are mostly based on target-driven
approaches, where an efficient compound to influence disease-
related molecular target is sought. The network approach
examines the effects of drugs in the context of cellular
networks. In this model, a drug-induced inhibition of a single
target means that the interactions around a given target are
eliminated, whereas partial inhibition can be modeled as a
partial knockout of the interactions of the target.
Cellular networks are usually damaged by random failures,
such as the oxidative damage of free radicals, the indirect effect
of somatic mutations and the complex phenomenon of
ageing [50] . Opposed to this, drug-driven network attacks are
targeted to find the most efficient way to influence network
behavior. Several classes of drugs, such as antibiotics, fungicides,
anticancer drugs as well as numerous chemical compounds,
such as pesticides, are designed to destroy the normal function
of cellular networks.
Networks have a number of vulnerable points and, therefore,
can be attacked in many ways ( Figure 3 ). The first major insight
identifying a set of weak points in natural networks may came
from the work of A L Barabasi and co-workers [30,51] , who
discovered that many real-world networks (including cellular
networks) have a scale-free degree distribution, which means
Table 1. Cellular networks as drug target maps.
Name of cellular
network
Network elements Network links Potential drug targets
Protein interaction
network
Cellular proteins Transient or permanent bonds Hubs, bridges, proteins in
modular centers and overlaps
Cytoskeletal network Cytoskeletal fi laments Transient or permanent bonds Cross-linking proteins
Organelle network Membrane segments
(membrane vesicles, domains
and of rafts of cellular
membranes) and cellular
organelles (mitochondria,
lysosomes, segments of the
endoplasmic reticulum, etc.)
Proteins, protein complexes
and/or membrane vesicles,
channels
Proteins and lipid rafts
regulating inter-organellar
junctions
Signaling network Proteins, protein complexes,
RNA (such as micro-RNA)
Highly specifi c interactions
undergoing a profound change
(either activation or inhibition)
when a specifi c signal reaches
the cell
Hubs, bridge, proteins of
cross-talks, ‘critical nodes’
Metabolic network Metabolites and small molecules,
such as glucose, or adenine, etc.
Enzyme reactions transforming
one metabolite to the other
Metabolic switch enzymes
and their regulatory proteins,
channelling
Gene transcription
network
Transcriptional factors or their
complexes and DNA gene
sequences
Functional (and physical)
interactions between transcription
factor proteins (sometimes RNAs)
and various parts of the gene
sequences in the cellular DNA
Hubs, bridges, proteins in
modular centers and overlaps

Korcsmáros, Szalay, Böde, Kovács & Csermely
Expert Opin. Drug Discov. (2007) 2(6) 5
A. Attacking hubs
C. Attacking bridges D. Attacking weighted bridges
B. Attacking hub-links
Figure 3. Attack scenarios on networks. In this fi gure the authors summarize a number of malicious attacks on vulnerable points
of networks. A. Attacks on nodes with highest degree (hubs). B. Attacks on ‘hub-links’ with the highest degree of their end points.
C. Attacks on bridging elements with links having a high betweenness centrality. D. Attacks on bridges having links with the highest
weighted centrality.
(in simple terms) that these networks have hubs (e.g., nodes
with a much greater number of neighbors than average). If hubs
are selectively attacked, the information transfer in most
scale-free networks soon becomes significantly hindered. In
other words: hubs are central points of networks. Scale-free
networks are, however, highly resistant against random damage.
These two features can be summarized as having a robust, but
fragile nature of scale-free networks. As an illustration for the
robustness, 99% of the internet in a random-attack strategy
may be deleted and the continuity of the network still remains,
that is, the internet could still be used after such an attack,
albeit it would be much slower than usual [52] . On the contrary,
malicious attacks on hubs may follow a ‘greedy’ strategy, meaning
that degrees are continuously recalculated after each attack and
network elements are re-ranked. This greedy strategy is often
more powerful than using the original degrees of network elements
throughout the whole process [53] .
Hubs are the centers of networks only from the point of
local network topology. Another approach for pinpointing
central elements of network communication is to find those
elements (or links) that are in a centered position, not in the
local, but in the global topology. The “betweenness of centrality”
of a link refers to the number of shortest paths between any
two elements of the network across the given link. The
betweenness centrality was worked out initially in social
networks [54] , but later it became a preferred centrality measure
to assess the presumed effect of targeted attacks on network
stability. Inverse geodesic length (also called network efficiency),
meaning the sum of the inverse of the shortest paths between
network elements, is widely used as an indicator of network
damage after the removal of links or elements [53,55,56] .
Alternative measures of the damaging element or link removal
have also been worked out by Latora and Marchiori [57] ,
which are based on monitoring of the performance of the
whole network.
Recent studies take into account the weights and directness
of network links. This is much closer to the real, cellular
scenario, where protein–protein interactions are characterized
by their affinity and/or prevalence (link weight) as well as
direction (e.g., in form of signaling). The removal of the
links with the highest weighted centralities is often more
devastating to network behavior than the removal of the
most central links based on the unweighted version of the same
network topology [58] .
Another recent approach is to take into account mesoscopic
centrality network topology measures, which are neither based

How to design multi-target drugs: target search options in cellular networks
6 Expert Opin. Drug Discov. (2007) 2(6)
on local information (such as hubs) or global information
(such as betweenness centrality), on network structure. Motter
et al. [59] found that the removal of long-range links connecting
elements a long distance from each other has a profound
impact on small-world networks; however, it fails to affect
many scale-free networks. Another complication is caused by
multi-layered networks [101] , where various communities and
modules are organized in a hierarchical fashion. These studies
reveal that the use of the complex structural information of
real world networks needs more sophisticated methods, such as
an integrated assessment of link density and network topology
(Kovacs et al. , manuscript in preparation).
The network approach helped to gain ground in drug target
analysis. Flux-balance analysis (or metabolic control analysis)
uses a large database of experimental data and calculates all
of the metabolic rates of the metabolic network, assuming
that the rates of reactions producing a metabolite must be
equal with the rates of reactions that consume it. Flux-balance
analysis of metabolic networks uncovers vulnerable points of
parasite or pathologic metabolisms providing potential targets
for efficient drug action [39,60-62] . Comparison of cellular
(transcriptional, signaling, protein–protein interaction etc.)
networks from various genomes helps to identify the function
of novel proteins and, thus, increases the number of potential
drug targets [63,64] . Analysis of protein–protein interactions
identifies protein contact surfaces as potential sites of drug
action [65] and neural networks have long been applied to help
in methodological and computational drug design [66] .
Most of the above network-related methods have been used
so far to steer target identification attempts to single targets
and a systematic network-based analysis of multi-target drug
action is still to come. In the authors’ earlier study [55,67] , a
multi-target attack on the genetic regulatory networks of the
bacterium Escherichia coli or the yeast Saccharomyces cerevisiae
was modeled. A comparison of various strategies suggested
that multiple, but partial, attacks on carefully selected targets
were almost inevitably more efficient than the knockout of a
single target. For example, the largest damage to the E. coli
regulatory network was reached by removing an element with
72 connections. However, the same damage could be achieved
if three to five nodes were partially inactivated. Multiple attacks
have proven to be more efficient than a single attack even if the
number of affected interactions remained the same [55,67] .
Thus, the reason underlying the efficiency of multi-target
attacks was proven not to be trivial, even from a theoretical
point of view: multi-target attacks were not only better because
they affected the network at more sites, but they could, espe-
cially if distributed in the entire network, perturb complex
systems more than concentrated attacks even if the number of
targeted interactions remained the same.
6. Network diseases
The authors’ initial network-based analysis of the potential
efficiency of multi-target drugs [55,67] was based on the topology
of bacterial and yeast gene regulatory networks, which may be
regarded as an initial model of the multi-target action of anti-
biotics and fungicides, where network damage corresponds
well to the desired drug action. For the analysis of multi-target
drugs affecting specific disease models (e.g., antihypertensive,
antipsychotic and antidiabetic drugs), more specific signaling,
metabolic and transcriptional network models are needed.
As a prelude of this process, several complex multifactorial
diseases have already been described as ‘network diseases’.
Cancer was assessed as a systems biology disease by Hornberg
et al. [68] . The complexity of intra- and extra-cellular cancer-
specific changes in signaling, gene-regulatory (and, most
probably, protein–protein interaction) networks, the profound
reorganization of cellular metabolism, the multiple types and
interactions of cells involved and the complexity of all these
events at various types and subtypes of malignant transformation
make the name ‘systems biology disease’ well deserved for all
stages of tumor development.
Network effects of continuously changing functional neuron
assemblies may provide an explanation of the daily fluctuations
in the symptoms of neurodegenerative diseases. This approach
may show novel pathways of drug development, leading to
shorter and cheaper clinical trials that concentrate on the
short-term attenuation of symptom fluctuations, instead of
waiting for and monitoring the long-term and rather elusive
benefits of inhibited neurodegeneration [69] . As increased
fluctuations of this efficiency may reflect a general decline in
the stability of the overall network, which is related to the
reduction of weak links [29,46] – multi-target drugs might be an
ideal choice to prevent further functional loss in this sense.
The complexity of neuronal networks also led to the concept
of network disease in the case of depression aiding the design
of novel, multi-target antidepressants [7] .
Last, but by no means least, ageing has also been conceived
as a network disease [50,70] . The multiple reasons and stages of
biological ageing, the increasing variability of symptoms and
malfunctions, both from one elderly person to the other and
from one day to another, all call for a network-based analysis
of the increasing amount of data collected so far.
7. Target-sets of multi-target drugs:
the help of networks
How should we find the relevant target-sets of multi-
target drugs? In the last decade several experimental and
modeling approaches have been developed to identify
single targets in a network context [39,60,61,71-74] . Appropriate
modifications of these approaches may constitute the first
step for zooming in to a smaller set of potential targets. A high-
throughput combinative screen of all possible combinations
may be a daunting task and prospect [5] . How can we pin
point those target combinations that might be relevant in the
clinical setting?
Surprisingly, old-fashioned drug development might come
back here to help: in vivo pharmacology (i.e., whole-animal

Korcsmáros, Szalay, Böde, Kovács & Csermely
Expert Opin. Drug Discov. (2007) 2(6) 7
studies) might become important again [75] . However, for more
efficient in vivo testing, better animal models are needed.
Better animal models can be achieved by humanizing the
metabolic and signaling network of test animals.
Can the network approach help suggest potential target
sets? The answer is not known at present. However, the authors
have many promising tools to assess the relevance of their
present, network-based knowledge on the complexity of the
cell in pathologic states. The authors will list these in the
Expert Opinion section.
8. Conclusion
Combinatorial therapies have recently become one of the most
successful drug development strategies. Multi-target drugs
can be perceived as siblings of combinatorial therapies, where
the different agents (often as many as five to six of them) are
compressed into a single, integrated chemical entity. Multi-
target drugs expand the number of pharmacologically relevant
target molecules by introducing a set of indirect, network-
dependent effects. Multi-target drugs usually have a low affinity
towards their targets. An increasing amount of evidence shows
that low affinity, especially if multiplied, does not mean low
efficiency. On the contrary, several signaling and regulatory
events are actually based on low-affinity interactions. Moreover,
low-affinity binding of multi-target drugs eases the constraints
of druggability and significantly increases the size of the
druggable proteome. These effects tremendously expand the
number of potential drug targets and will introduce novel
classes of multi-target drugs with smaller side effects and
toxicity. Cellular networks offer a large number of possibilities,
such as hubs and bridges with high betweenness centrality
to find target-sets of multi-target drugs. However, the ‘fuzzy-
ness’ (meaning uncertain and incomplete information) of
cellular networks and the data sets on network diseases,
such as cancer, diabetes and neurodegeneration, need a more
sophisticated approach.
9. Expert opinion: network-based, smart
multi-target drugs of the future
We predict that in 5 – 10 years multi-target drugs will be
much more common than it is today. The emerging knowl-
edge of recent years strongly suggests that these drugs have
a better chance of affecting the complex equilibrium of
whole cellular networks than drugs acting on a single
target. Target expansion to indirect targets will lead to the
dis covery of several novel classes of drugs. Moreover, it is
sufficient that these multi-target drugs affect their targets
only partially, which multiplies our target choices from the
point of druggability.
Low-affinity, multi-target drugs might have another
advantage. Weak links have been shown to stabilize complex net-
works, including macromolecular networks, ecosystems and social
networks, buffering the changes after system perturbations.
If multi-target, low-affinity drugs inhibit their targets, they
change a strong link into a weak link, instead of eliminating
the link completely. A weak activation also results in a weak
link in most of the cases. Thus, multi-target drugs can increase
the number of weak links in cellular networks and, thus,
stabilize these networks, in addition to having multiple
effects. Stabilization of the cells becomes especially important
if we take into account that cells of stressed, sick and ageing
organisms are at the ‘verge of chaos’, showing a much greater
instability than their healthy counterparts [29,46,70,76,77] . Thus,
multi-target drugs may have multiple beneficial effects:
• they can be designed to act on a carefully selected set of
primary targets, which help to sum up the action of the
drug on key, therapeutically relevant secondary targets
• via the above-mentioned, indirect approach and their low
affinity binding multi-target drugs may avoid the presently
common dual-trap of drug resistance and toxicity
• finally, the low affinity, weak links of multi-target drugs
have an important side effect: they stabilize the sick cell,
which may be sometimes at least as beneficial as their
primary therapeutic effect.
What is the 5 – 10 year perspective on fi nding the relevant
target-sets of multi-target drugs? Here, our increasing knowledge
of cellular networks will certainly play a key role in the future.
The increasing complexity of the data sets requires more
sophisticated tools to direct our attention to relevant areas of
the information universe. In the last part of this review, we
list a few ideas for future directions in the development of
network analysis tool-kits:
• overlaps of network communities [31] , which have a key role
in the regulation of complex systems [29] , may be efficient
guides to restrict the initial target pool in search of target
sets for multi-target drugs
• differential analysis of relevant cellular networks, (such as
signaling networks, gene-regulatory networks, metabolic net-
works and protein–protein interaction networks) of healthy
and sick cells may provide an even more efficient screen
• finally, the present development of analytical tools to assess
the evolution and dynamism of whole cellular networks
will reveal search methods to deconvolute the hidden
masterminds of the primary target sets from the presently
known, pharmacologically relevant secondary targets
(Kovacs et al. , patent application submitted). A variant of
this approach has been called ‘reverse-engineering’, which
deciphers potential targets by the analysis of the effect of
a limited set of experimental perturbations [71,78-80] . These
approaches open the way to finding multiple-targets and
to designing alternative target sets to mimic the action of
existing, successful drugs
In summary, 10 years from now we will have a multitude
of expanded cellular data sets having detailed and differential
information on variable pairs of healthy and sick cells. These
data sets will have graded information, which enables the

How to design multi-target drugs: target search options in cellular networks
8 Expert Opin. Drug Discov. (2007) 2(6)
building of weighted and directed networks. Analytical tools
will be developed to assess the complexity of these networks,
keeping all data and giving a zoom in picture of any resolution
in a computationally accessible, short time. Multi-target
drugs will magnify the presently available target fi eld by
introducing thousands of secondary targets, as well as other
thousands of druggable proteins. This will all lead to the
discovery of several entirely novel drug classes. Finally, we will
enjoy the network option of fi ne-tuning of drug action of
multi-target drugs by the targeted manipulation of certain
elements of the already existing target sets of multi-target
drugs, which may lead to the development of personalized
medicines – at a low cost.
Acknowledgements
The authors would like to thank the three anonymous referees
for their helpful comments. Supported in part by funds from
the EU (FP6-506850, FP6-016003) and the Hungarian Office
of Research and Development (NFKP-1A/056/2004).
Bibliography
Papers of special note have been highlighted
as either of interest (•) or of considerable
interest (••) to readers.
1. BROWN D, SUPERTI-FURGA G:
Rediscovering the sweet spot in drug
discovery. Drug Discov. Today (2003)
8:1067-1077.
• An important review on target-search
options.
2. OVERINGTON JP, AL-LAZIKANI B,
HOPKINS AL: How many drug targets
are there? Nat. Rev. Drug Discov. (2006)
5:993-996.
3. SZUROMI P, VINSON V,
MARSHALL E: Rethinking drug
discovery. Science (2004) 303:1795.
4. BORISY AA, ELLIOTT PJ,
HURST NW et al.: Systematic discovery
of multicomponent therapeutics.
Proc. Natl. Acad. Sci. USA (2003)
100:7977-7982.
• A key paper on combination therapy-
based drug design.
5. DANCEY JE, CHEN HX: Strategies for
optimizing combinations of molecularly
targeted anticancer agents. Nat. Rev.
Drug Discov. (2006) 5:649-659.
6. KEITH CT, BORISY AA,
STOCKWELL BR: Multicomponent
therapeutics for networked systems.
Nat. Rev. Drug Discov. (2005)
4:71-78.
7. MILLAN MJ: Multi-target strategies
for the improved treatment of
depressive states: conceptual foundations
and neuronal substrates, drug discovery
and therapeutic application.
Pharmacol. Therap. (2006) 110:135-370.
8. ZIMMERMANN GR, LEHAR J,
KEITH CT: Multi-target therapeutics:
when the whole is greater than the sum
of the parts. Drug Discov. Today (2007)
12:34-42.
9. FITTER S, JAMES R: Deconvolution
of a complex target using DNA aptamers.
J. Biol. Chem. (2005) 280:34193-34201.
10. ROLLINGER JM, LANGER T,
STUPPNER H: Strategies for effi cient lead
structure discovery from natural products.
Curr. Med. Chem. (2006) 13:1491-1507.
11. SHAFIKHANI SH, ENGEL J:
Pseudomonas aeruginosa type III-secreted
toxin ExoT inhibits host-cell division by
targeting cytokinesis at multiple steps.
Proc. Natl. Acad. Sci. USA (2006)
103:15605-15610.
12. PAPP B, PAL C, HURST LD: Metabolic
network analysis of the causes and
evolution of enzyme dispensability in yeast.
Nature (2004) 429:661-664.
13. PAL C, PAPP B, LERCHER MJ,
CSERMELY P, OLIVER SG, HURST LD:
Chance and necessity in the evolution
of minimal metabolic networks. Nature
(2006) 440:667-670.
14. KITANO H: A robustness-based approach
to systems-oriented drug design. Nat. Rev.
Drug Discov. (2007) 6:202-210.
15. SHAROM JR, BELLOWS DS,
TYERS M: From large networks to
small molecules. Curr. Op. Chem. Biol.
(2004) 8:81-90.
16. FRANTZ S: Drug discovery: playing
dirty. Nature (2005) 437:942-943.
17. HUANG S: Rational drug discovery:
what can we learn from regulatory
networks? Drug Discov. Today (2002)
7:S163-S169.
18. KAELIN WG JR: Gleevec: prototype
or outlier? Science STKE (2004)
(225), pe12.
19. MORPHY R, KAY C, RANKOVIC Z:
From magic bullets to designed multiple
ligands. Drug Discov. Today (2004)
9:641-651.
• A structural analysis of multi-target
drug design.
20. ROGAWSKI MA: Low affi nity
channel blocking (uncompetitive)
NMDA receptor antagonists as
therapeutic agents – towards an
understanding of their favorable
tolerability. Amino Acids (2000)
19:133-149.
21. SATHORNSUMETEE S,
VREDENBURGH KA,
LATTIMORE KP, RICH JN: Malignant
glioma drug discovery – targeting protein
kinases. Expert Opin. Drug Discov.
(2007) 2:1-17.
22. SOTI C, NAGY E, GIRICZ Z, VIGH L,
CSERMELY P, FERDINANDY P: Heat
shock proteins as emerging therapeutic
targets. Br. J. Pharmacol. (2005)
146:769-780.
23. SREEDHAR AS, SOTI C, CSERMELY P:
Inhibition of Hsp90: a new strategy for
inhibiting protein kinases. Biochim.
Biophys. Acta (2004) 1697:233-242.
24. YOUDIM MBH, BUCCAFUSCO JJ:
Multi-functional drugs for various CNS
targets in the treatment of
neurodegenerative disorders. Trends
Pharmacol. Sci. (2005) 26:27-36.
25. SHARON J, LIEBMAN MA,
WILLIAMS BR: Recombinant
polyclonal antibodies for cancer therapy.
J. Cell Biochem. (2005) 96:305-313.
26. TODOROVSKA A, ROOVERS RC,
DOLEZAL O, KORTT AA,
HOOGENBOOM HR, HUDSON PJ:
Design and application of diabodies,
triabodies and tetrabodies for cancer
targeting. J. Immunol. Methods (2001)
248:47-66.
27. BARABASI A-L, OLTVAI ZN: Network
biology: understanding the cell’s functional
organization. Nat. Rev. Genet. (2004)
5:101-113.
• A comprehensive and brief summary
of the properties of biological
networks.

Korcsmáros, Szalay, Böde, Kovács & Csermely
Expert Opin. Drug Discov. (2007) 2(6) 9
28. BOCCALETTI S, LATORA V,
MORENO Y, CHAVEZ M,
HWANG D-U: Complex networks:
structure and dynamics. Physics Rep.
(2006) 424:175-308.
29. CSERMELY P: Weak Links: A Universal
Key for Network Diversity and Stability.
Springer Verlag, Heidelberg (2006).
• A comprehensive and easily
understandable summary of network
science from a biological perspective
with more than 800 original references.
30. BARABASI A-L, ALBERT R:
Emergence of scaling in random
networks. Science (1999) 286:509-512.
• A key paper on the generality of scale-free
degree distribution of networks.
31. PALLA G, DERENYI I, FARKAS T,
VICSEK T: Uncovering the overlapping
community structure of complex
networks in nature and society. Nature
(2005) 435:814-818.
• The paper describes a method to fi nd
overlapping network modules.
32. WATTS DJ, STOGATZ SH: Collective
dynamics of ‘small-world’ networks.
Nature (1998) 393:440-442.
• A key paper on the generality of the
small-world character of networks.
33. ARITA M: The metabolic world of
Escherichia coli is not small. Proc. Natl.
Acad. Sci. USA (2004) 101:1543-1547.
34. TANAKA R, YI TM, DOYLE J: Some
protein interaction data do not exhibit
power law statistics. FEBS Lett. (2005)
579:5140-5144.
35. CHAUDHURI A, CHANT J:
Protein-interaction mapping in search
of effective drug targets. Bioessays (2005)
27:958-969.
36. TANIGUCHI CM, EMANUELLI B,
KAHN CR: Critical nodes in signalling
pathways: insights into insulin action.
Nat. Rev. Mol. Cell Biol. (2006) 7:85-96.
37. SANTONICO E, CASTAGNOLI L,
CESARENI G: Methods to reveal
domain networks. Drug Discov. Today
(2005) 10:1111-1117.
38. WILSON D, MADERA M,
VOGEL C, CHOTHIA C, GOUGH J:
The SUPERFAMILY database in 2007:
families and functions. Nucleic Acids Res.
(2007) 35:D308-D313.
39. KELL DB: Systems biology, metabolic
modelling and metabolomics in drug
discovery and development. Drug Discov.
Today (2006) 11:1085-1092.
40. MICHAELIS ML, SEYB KI, ANSAR S:
Cytoskeletal integrity as a drug target.
Curr. Alzheimer Res. (2005) 2:227-229.
41. VON MERING C, KRAUSE R, SNEL B
et al.: Comparative assessment of large-scale
data sets of protein-protein interactions.
Nature (2002) 417:399-403.
42. SHEN J, ZHANG J, LUO X et al.:
Predicting protein-protein interactions
based only on sequences information.
Proc. Natl. Acad. Sci. USA (2007)
104:4337-4341.
43. KIM PM, LU LJ, XIA Y, GERSTEIN MB:
Relating three-dimensional structures to
protein networks provides evolutionary
insights. Science (2006) 314:1938-1941.
44. LIPTON SA: Turning down, but not
off. Neuroprotection requires a paradigm
shift in drug development. Nature (2004)
428:473.
45. CHENG Y, LEGALL T, OLDFIELD CJ
et al.: Rational drug design via intrinsically
disordered protein. Trends Biotechnol.
(2006) 24:435-442.
46. CSERMELY P: Strong links are
important, but weak links stabilize
them. Trends Biochem. Sci. (2004)
29:331-334.
47. AMES BN: Low micronutrient
intake may accelerate the degenerative
diseases of aging through allocation
of scarce micronutrients by triage.
Proc. Natl. Acad. Sci. USA (2006)
103:17589-17594.
48. KLEIN P, PAWSON T, TYERS M:
Mathematical modeling suggests
cooperative interactions between a
disordered polyvalent ligand and a
single receptor site. Curr. Biol. (2003)
13:1669-1678.
49. NASH P, TANG X, ORLICKY S et al.:
Multisite phosphorylation of a CDK
inhibitor sets a threshold for the onset
of DNA replication. Nature (2001)
414:514-521.
50. KIRKWOOD TBL, KOWALD A:
The network theory of aging.
Exp. Gerontol. (1997) 32:395-399.
51. ALBERT R, JEONG H, BARABASI A-L:
Attack and error tolerance of complex
networks. Nature (2000) 406:378-382.
• A key paper on the vulnerability of
networks on random and targeted attacks.
52. COHEN R, EREZ K,
BEN-AVRAHAM D, HAVLIN S:
Resilience of the Internet to random
breakdowns. Phys. Rev. Lett. (2000)
85:4626-4628.
53. HOLME P, KIM BJ, YOON CN,
HAN SK: Attack vulnerability of
complex networks. Phys. Rev. E (2002)
65:056109.
54. FREEMAN LC: A set of measuring
centrality based on betweenness.
Sociometry (1977) 40:35-41.
55. AGOSTON V, CSERMELY P,
PONGOR S: Multiple hits confuse
complex systems: a genetic network as
an example. Phys. Rev. E (2005)
71:051909.
• An important early model of the
potential effect of multi-target drugs
on cellular networks.
56. LATORA V, MARCHIORI M:
Effi cient behavior of small-world
networks. Phys. Rev. Lett. (2001)
87:198701.
57. LATORA V, MARCHIORI M:
Vulnerability and protection of
infrastructure networks. Phys. Rev. E
(2005) 71:015103.
58. DALL’ASTA L, BARRAT A,
BARTHÉLEMY M, VESPIGNANI A:
Vulnerability of weighted networks.
J. Stat. Mech. (2006) P04006.
59. MOTTER AE, NISHIKAWA T, LAI Y:
Range-based attack on links in scale-free
networks: are long-range links responsible
for the small-world phenomenon?
Phys. Rev. E (2002) 66:065103.
60. CASCANTE M, BOROS LG,
COMIN-ANDUIX B, DE ATAURI P,
CENTELLES JJ, LEE PWN: Metabolic
control analysis in drug discovery and
disease. Nat. Biotechnol. (2002)
20:243-249.
61. CORNISH-BOWDEN A,
CARDENAS ML: Metabolic analysis
in drug design. C. R. Biol. (2003)
326:509-515.
62. DUARTE NC, BECKER SA,
JAMSHIDI N et al.: Global reconstruction
of the human metabolic network
based on genomic and bibliomic data.
Proc. Natl. Acad. Sci. USA (2007)
104:1777-1782.
63. BERGMANN S, IHMELS J, BARKAI N:
Similarities and differences in genome-wide
expression data of six organisms.
PLoS Biol. (2004) 2:85-93 (E9).
64. STUART JM, SEGAL E, KOLLER D,
KIM SK: A gene-coexpression network

How to design multi-target drugs: target search options in cellular networks
10 Expert Opin. Drug Discov. (2007) 2(6)
for global discovery of conserved genetic
modules. Science (2003) 302:249-255.
65. ARCHAKOV AI, GOVORUN VM,
DUBANOV AV et al.: Protein-protein
interactions as a target for drugs in
proteomics. Proteomics (2003) 3:380-391.
66. AOYAMA T, SUZUKI Y, ICHIKAWA H:
Neural networks applied to quantitative
structure-activity relationship analysis.
J. Med. Chem. (1990) 33:2583-2590.
67. CSERMELY P, AGOSTON V,
PONGOR S: The effi ciency of multi-target
drugs: the network approach might help
drug design. Trends Pharmacol. Sci. (2005)
26:187-182.
•• This paper fi rst introduces the concept of
network-based analysis in the action of
multi-target drugs and lists several lines
of evidence to show that low affi nity,
multi-target drugs may perturb networks
more effi ciently than high-affi nity,
single-hit drugs.
68. HORNBERG JJ, BRUGGEMAN FJ,
WESTERHOFF HV, LANKELMA J:
Cancer: a systems biology disease.
Biosystems (2006) 83:81-90.
69. PALOP JJ, CHIN J, MUCKE L:
A network dysfunction perspective on
neurodegenerative diseases. Nature
(2006) 443:768-773.
70. CSERMELY P, SOTI C: Aging cellular
networks: chaperones as major participants.
Exp. Gerontol. (2007) 42:113-119.
71. GARDNER TS, DI BERNARDO D,
LORENZ D, COLLINS JJ: Inferring
genetic networks and identifying
compound mode of action via expression
profi ling. Science (2003) 301:102-105.
• An important example of dynamic
network modelling of drug action.
72. GIAEVER G, FLAHERTY P, KUMM J
et al.: Chemogenomic profi ling:
identifying the functional interactions
of small molecules in yeast. Proc. Natl.
Acad. Sci. USA (2004) 101:793-798.
73. LUM PY, ARMOUR CD,
STEPANIATS SB et al.: Discovering
modes of action for therapeutic
compounds using a genome-wide screen
of yeast heterozygotes. Cell (2004)
116:121-137.
74. PARSONS AB, LOPEZ A, GIVONI IE
et al.: Exploring the mode-of-action of
bioactive compounds by chemical-genetic
profi ling in yeast. Cell (2006) 126:611-625.
75. IN VIVO PHARMACOLOGY
TRAINING GROUP: The fall and rise of
in vivo pharmacology. Trends Pharmacol.
(2002) 23:13-18.
76. GOLDBERGER AL, AMARAL LAN,
HAUSDORF JM, IVANOV PC,
PENG C-K, STANLEY HE: Fractal
dynamics in physiology: alterations
with disease and aging. Proc. Natl. Acad.
Sci. USA (2002) 99:2466-2472.
77. SZALAY M, KOVACS IA,
KORCSMAROS T, BODE C,
CSERMELY P: Stress-induced
rearrangements of cellular networks:
consequences for protection and drug
design. FEBS Lett. (in press),
arxiv.org/q-bio.MN/0702006.
78. HALLEN K, BJORKEGREN J,
TEGNER J: Detection of compound
mode of action by computational
integration of whole-genome measurements
and genetic perturbations. BMC
Bioinformatics (2006) 7:51.
•• A careful analysis of network dynamics
to assess drug action.
79. KHOLODENKO BN,
KIYATKIN A, BRUGGEMAN FJ,
SONTAG E, WESTERHOFF HV,
HOEK JB: Untangling the wires: a strategy
to trace functional interactions in signaling
and gene networks. Proc. Natl. Acad.
Sci. USA (2002) 99:12841-12846.
• An important paper on the dynamical
analysis of networks.
80. TEGNÉR J, YEUNG MK, HASTY J,
COLLINS JJ: Reverse engineering gene
networks: integrating genetic perturbations
with dynamical modeling. Proc. Natl. Acad.
Sci. USA (2003) 100:5944-5949.
• An important paper on the dynamical
analysis of networks.
Website
101. www.arxiv.org/physics/0610018
KURANT M, THIRAN P: Error and
attack tolerance of layered complex
networks (2006).
Affi liation
Tamás Korcsmáros1, Máté S Szalay1,
Csaba Böde2 PhD, István A Kovács1 &
Péter Csermely†1 PhD, DSc
†Author for correspondence
1Semmelweis University, Department of Medical
Chemistry, PO Box 260, H-1444 Budapest 8,
Hungary and Predinet Ltd., Dongo Street 8,
H-1149 Budapest, Hungary
Tel: +36 1 266 2755; Fax: +36 1 266 6550;
E-mail: csermely@predinet.com
2Semmelweis University, Department of
Biophysics and Radiation Biology, PO Box 263,
H-1444 Budapest 8, Hungary