A guide to drug discoveryHit and lead generation: beyond high-throughput screening

© 2003 Nature Publishing Group

R E V I E W S

NATURE REVIEWS | DRUG DISCOVERY VOLUME 2 | MAY 2003 | 369

All drugs that are presently on the market are estimatedto target less than 500 biomolecules, ranging fromnucleic acids to enzymes, G-protein-coupled receptors(GPCRs) and ion channels1 (FIG. 1). Although the targetportfolio of large pharmaceutical companies is continu-ously changing, all of the main target classes are likely tobe represented. The relative distribution of classes variesfrom company to company on the basis of the diseasearea they focus on, and also because some target familiesare more numerous than others. Presently, GPCRs arethe predominant target family addressed, and morethan 600 genes encoding GPCRs have been identifiedfrom human genome sequencing efforts2.

The balance of such targets and their relative noveltyare the domain of the companies’ early project-portfoliomanagement strategy3. Of course, a balance always hasto be struck between the requirements of the diseasearea for efficacious new therapies, business considera-tions and, most crucially, the chemical tractability orDRUGABILITY of targets for small-molecule intervention4.It is well accepted within the medicinal chemistry com-munity that, independently of the technology applied,certain protein families are more readily modulated bysmall-molecule intervention than others. In this context,

target selection plays a pivotal role in the final outcomeof HIT and LEAD identification activities. A retrospectiveanalysis of past discovery programmes reveals thatmuch higher success rates have been demonstrated foraminergic GPCRs compared with large peptide recep-tors, for example. This is not surprising, as modulatingprotein–protein interactions — often involving largesurface areas — by a small chemical entity is far moredemanding than competing against an endogenoussmall-molecule ligand.

Apart from the intrinsic biochemical and kineticchallenges in identifying an appropriate modulator fora target, the range of meaningful assays and ligand-identification technologies can also significantly influ-ence the chances of success. Considering a representativetarget portfolio, HIGH-THROUGHPUT SCREENING (HTS) ispresently the most widely applicable technologydelivering chemistry entry points for drug discoveryprogrammes. However, it is well recognized that evenwhen compounds are identified from HTS they are notalways suitable for the initiation of further medicinalchemistry exploration (FIG. 2). The potential for successis nevertheless demonstrated by a variety of develop-ment candidates and marketed drugs that have resulted

HIT AND LEAD GENERATION:BEYOND HIGH-THROUGHPUTSCREENINGKonrad H. Bleicher, Hans-Joachim Böhm, Klaus Müller and Alexander I. Alanine

The identification of small-molecule modulators of protein function, and the process of transformingthese into high-content lead series, are key activities in modern drug discovery. The decisions takenduring this process have far-reaching consequences for success later in lead optimization and evenmore crucially in clinical development. Recently, there has been an increased focus on theseactivities due to escalating downstream costs resulting from high clinical failure rates. In addition,the vast emerging opportunities from efforts in functional genomics and proteomics demands adeparture from the linear process of identification, evaluation and refinement activities towards amore integrated parallel process. This calls for flexible, fast and cost-effective strategies to meet thedemands of producing high-content lead series with improved prospects for clinical success.

F. Hoffmann-La Roche Ltd,Grenzacherstrasse 124,CH-4070, Basel, Switzerland.Correspondence to A. A.e-mail: [email protected]:10.1038/nrd1086

DRUGABILITY

The feasibility of a target to beeffectively modulated by a smallmolecule ligand that hasappropriate bio-physico-chemical and absorption,distribution, metabolism andexcretion properties to bedeveloped into a drug candidatewith appropriate properties for the desired therapeutic use.

A G U I D E TO D R U G D I S C O V E R Y


HIT

A primary active compound(s),with non-promiscuous bindingbehaviour, exceeding a certainthreshold value in a givenassay(s). The ‘active’ is followedup with an identity and purityevaluation, an authentic sample isthen obtained or re-synthesizedand activity confirmed in amulti-point activitydetermination to establish thevalidity of the hit (validated hit).

370 | MAY 2003 | VOLUME 2 www.nature.com/reviews/drugdisc

R E V I E W S

‘hits’, are progressed into lead series by a comprehensiveassessment of chemical integrity, synthetic accessibility,functional behaviour, STRUCTURE–ACTIVITY-RELATIONSHIPS

(SAR), as well as bio-PHYSICOCHEMICAL and absorption,distribution, metabolism and excretion (ADME)properties. This early awareness of the required profile(a given selectivity, solubility, permeation, metabolicstability and so on) is important for the selection andprioritization of series with the best development poten-tial. In this regard, it is important that at least two leadseries of significantly different pharmacological and/orstructural profile are advanced as reserve, or ‘back-up’,lead series. This insures against unexpected failures dueto unpredictable factors, such as toxicological findings inlater animal studies. The effect of such a rigorousprocess at an early stage is to achieve greater awarenessof key liabilities, which can be addressed in adequatetime and with sufficient resources. The net effect is toreduce attrition in the costly clinical phases by intercept-ing many crucial ADME-related issues before they arediscovered too late to be resolved.

Traditionally, hit identification is assumed to be thecrucial bottleneck for lead generation success, but this isnot the case. Rather, it is the overall characteristics of acompound class that make it an attractive starting pointfor medicinal chemists. Depending on the threshold set,an HTS campaign will always deliver active compounds,but it is the potential to optimize them into drug-like andinformation-rich lead series that is evidently far moreimportant for the downstream success of the entities.This is clearly illustrated by the observation that despitethe massive growth in screening compound numbersover the past 15–20 years, no corresponding increase insuccessfully launched new chemical entities has resulted.

Multi-property optimizationDuring the past few years, there has been an increas-ing awareness of the need for developing drug-likeproperties of a molecule. These are the balance of bio-physicochemical requirements for the molecule toreach its site of action in man at the given concentra-tion, for the necessary duration and with an adequatesafety window in order to answer the therapeuticprinciple hypothesis6.

In the past, lead-finding activities were mainly directedtowards affinity and selectivity rather than molecularproperties, metabolic liabilities and so on. It was notuncommon for a confirmed single primary active com-pound to be considered a ‘lead’ structure, or, in the case ofa cluster of actives with SAR, a ‘lead series’. Frequently,attention was not paid to characteristics of the moleculesother than perhaps their chemical stability and syntheticaccessibility.A consequence of these insufficient lead cri-teria — varying significantly not just between companiesbut also often within them — was that full project teamswere assembled with only a single superficially evaluated‘lead’.A thorough consideration of other important drugfeatures was often postponed until late in the optimiza-tion phase, when the in vitro affinity and selectivity hadbeen fully optimized at the expense of other facets, suchas solubility, permeability or metabolic stability.

from hits generated by HTS campaigns. It is evident thatin the future the overwhelming number of emergingtargets will dramatically increase the demands put onHTS and that this will call for new hit and lead generationstrategies to curb costs and enhance efficiency5.

Reducing attritionThe late-stage attrition of chemical entities in develop-ment and beyond is highly costly, and therefore suchfailures must be kept to a minimum by setting in place arigorous, objective quality assessment at key points inthe discovery process (FIG. 3). This assessment needs tobegin as early as possible and must be of high stringencyto prevent precious resources being squandered on lesspromising lead series and projects. The earliest point atwhich such knowledge-driven decisions can be made isin the lead-generation phase. Here, the initial actives, or

Nuclear receptors 2%

Ion channels 5%

Unknown 7%

Receptors 45% Enzymes 28%

Hormones and factors 11%

Nucleic acids 2%

Figure 1 | Therapeutic target classes. All current therapeutic targets can be subdivided intoseven main classes, wherein enzymes and receptors represent the largest part. Adapted withpermission from REF. 1 © American Association for the Advancement of Science (2000).

Figure 2 | Don’t panic… Turning an organic compound into a HIGH-CONTENT CHEMICAL LEAD

SERIES is a challenging and sometimes extremely complex endeavour, as numerous hurdlesbeyond activity and selectivity have to be overcome. It is vital to identify high-quality actives, or ‘hits’, as the molecular starting point is crucial in determining the later potential for success.Hit discovery and lead generation is therefore far more than just the identification of activecompounds; it is the multi-disciplinary process of selecting the most promising lead candidatesfrom rigorously assessed molecular series.



R E V I E W S

specifically acting low-molecular-weight modulatorswith an adequate activity in a suitable target assay. Suchinitial hits can be generated in a number of ways, depend-ing on the level of information available8. It is thereforeimportant to employ alternative hit-identificationstrategies that are able to tackle a variety of biologicalmacromolecular targets effectively, and to identify pro-prietary, synthetically tractable and pharmacologicallyrelevant compounds rapidly (FIG. 4).

These methods can be subdivided into those thatrequire very detailed ligand and/or target information,and those that do not. The former include techniquessuch as mutagenesis, NUCLEAR MAGNETIC RESONANCE (NMR)and X-ray crystallography, as well as the recognitioninformation that can be derived from endogenous lig-ands or non-natural small-molecule surrogates retrievedfrom literature and patents. At the other extreme are thetechnologies that do not require any prior informationon target or ligand, and which use serendipity-basedsearch strategies in either a given physical or virtual com-pound subset. Examples of so-called ‘random’ orpseudo-biased hit-identification strategies include bio-physical and biochemical testing that employ one orother method of detecting a molecular-binding event,usually in a high-throughput format9.

Between these extremes are more integratedapproaches, including targeted libraries and chemo-genomics10. The marriage of HTS with computationalchemistry methods11 has allowed a move away frompurely random-based testing, towards more meaning-ful and directed iterative rapid-feedback searches ofsubsets and focused libraries. The prerequisite for suc-cess of both approaches is the availability of the high-est-quality compounds possible for screening, eitherreal or virtual.

Quality versus quantityBesides the debate about how large a corporate com-pound collection should be, the questions of how tojudge the quality of the inventory, and how to ultimatelyimprove it, are important issues12. The collections of

Unfortunately, as the lead molecule becomes increas-ingly more potent, selective and tailored for the target,there is generally less tolerance for introducing significantchanges to affect biophysical properties without a largeintrinsic affinity penalty. Such unbalanced, sub-optimalcandidates entering clinical studies have attractive in vitroprofiles but poor ADME attributes that often precludethem from progressing and being fully evaluated in theclinic due to, for example, dose-limiting solubility, poorabsorption, CYTOCHROME P

450interactions or metabolic

instability. Clearly, poor initial leads with weak entrycriteria into lead optimization often can not be refined togenerate compounds with an appropriate profile, result-ing in high attrition rates at the clinical candidate selec-tion stage. This point has been highlighted in a recentanalysis of launched drugs, which indicates that, gener-ally, relatively minor changes in structural and physicalmolecular properties take place between the lead and thelaunched drug candidate7. This emphasizes once morethat the quality of the lead is crucial in most cases to thesuccess of the refinement and development process. If theclinical entry criteria are lax, the attrition is moved furtherinto pilot safety testing or early clinical-phase studies. Theoptimization process has historically been largelysequential in nature, addressing one issue at a time, withthe hope that all necessary modifications could beaccommodated within the PHARMACOPHORE optimized foraffinity only. This approach led to a very high andexpensive failure rate in the clinic for all major pharma-ceutical companies. During the mid-90s, this viewchanged to embrace a more holistic attitude towardslead optimization and subsequently to hit-to-lead gen-eration. The required trade-off for balancing theseproperties, in conjunction with pure affinity to achievean equilibrated potential therapeutic drug molecule,resulted in a change of approach from sequential toMULTI-DIMENSIONAL OPTIMIZATION.

Hit and lead generation strategiesThe entry point for any chemistry programme withindrug discovery research is generally the identification of

LEAD

A prototypical chemicalstructure or series of structuresthat demonstrate activity andselectivity in a pharmacologicalor biochemically relevant screen.This forms the basis for afocused medicinal chemistryeffort for lead optimization anddevelopment with the goal ofidentifying a clinical candidate.A distinct lead series has aunique core structure and theability to be patented separately.

HIGH-THROUGHPUT

SCREENING

Screening (of a compoundcollection) to identify hits in an in vitro assay, usuallyperformed robotically in 384-well microtitre plates.

HIGH-CONTENT LEAD SERIES

A lead series in which represen-tatives have been extensivelyrefined in not only theirstructure–activity relationshipand selectivity, but also in theirphysicochemical and earlyabsorption, distribution,metabolism and excretionproperties, and safety measures,such as metabolic stability,permeation and hERG liabilities.Correlations have been eluci-dated and all crucial parametershave shown themselves to bemodulated in the series.

STRUCTURE–ACTIVITY

RELATIONSHIP

The consistent correlation ofstructural features or groupswith the biological activity ofcompounds in a given biological assay.

PHYSICOCHEMICAL PROPERTIES

Physical molecular properties of a compound. Typicalproperties are solubility, acidity,lipophilicity, polar surface area,shape, flexibility and so on.

VALIDATED HIT SERIES

A set of hits clustered into sub-structurally related families,representatives of which havebeen evaluated for theirspecificity, selectivity,physicochemical and in vitroADME properties tocharacterize the series.

Hit generation

Hits

Lead generation

Leads

Lead optimization

Clinicalcandidate

Early knowledge;improved decision-making Reduce attrition rates

VHS LSI CCS

Target andhit identification

Hitrefinement

Leadrefinement

Regulatorydevelopment

$ $$ $$$ $$$$

Figure 3 | Stage-by-stage quality assessment to reduce costly late-stage attrition. Typical important milestones areVALIDATED HIT SERIES (VHS), LEAD SERIES IDENTIFIED (LSI) and clinical candidate selection (CCS), which ensure that only drugcandidates with an appropriately high-potential profile are advanced to the next phase.



R E V I E W S

starting points for a hit-to-lead programme is far morecomplex. Besides the variable perceptions of medicinalchemists of what makes a valuable hit (lead-like versusdrug-like), the issue concerning structural similarity,and in particular the overlap of chemical space (inven-tory versus vendor library), is frequently debated. Here,various computational algorithms are applied for thevalidation of compound collections to be purchased interms of their DRUG-LIKENESS13, chemical DIVERSITY14 andsimilarity to the existing corporate compoundinventory15. Although prediction tools for physicochem-ical properties, or ‘FREQUENT-HITTER’ LIABILITIES16, and so onare successfully applied in a routine fashion, the issueconcerning diversity is largely unresolved. The value ofstructure- or TOPOLOGY-oriented diversity DESCRIPTORS isnot in question, although determining pharmacologi-cally relevant similarity is far more complex and cannotbe described accurately by any single metric, such asbinding affinity. Conversely, it is difficult to describe thesimilarity (or dissimilarity) of two compounds thatdisplay the same activity, but possess, for example, dif-ferent functionality, selectivity, toxicological liabilitiesand so on. Similarity is a context-dependent parameterand therefore the context must define the appropriate

large pharmaceutical companies are approaching approx-imately one million entities, which represents historicalcollections (intermediates and precursors from earliermedicinal or agrochemical research programmes),natural products and COMBINATORIAL CHEMISTRY libraries.This is about an order of magnitude higher than tenyears ago when HTS and combinatorial chemistry firstemerged. Although this number is somewhat arbitrary,logistical hurdles and cost issues make this inventory sizean upper limit for most companies. Many research orga-nizations subsequently scaled back their large com-pound-production units after the realization that thequality component needed to get reliable and informa-tion-rich biological readouts cannot be obtained usingsuch ultra high-throughput synthesis technologiesfavoured in the early 1990s. Today, instead of huge inter-nal combinatorial chemistry programmes, purchasingefforts in every pharmaceutical company are directedtowards constantly improving and diversifying the com-pound collections, and making them globally availablefor random HTS campaigns.

Although the chemical integrity of compounds canbe checked by various analytical techniques, determin-ing whether the chemical entities are useful in general as

LEAD SERIES IDENTIFIED

A peer-reviewed milestone, therequirements to be fulfilled areclosely linked to the clinicalcandidate profile. Initial criteriaare defined when hits are firstidentified; they include activity,selectivity and pertinentphysicochemical properties, plusan evaluation of ADME andcertain safety attributes. In vivoactivity is not a mandatoryrequirement, provided theobstacles are appreciated andconsidered to be surmountablebased on evidence.

CYTOCHROME P450

A family of promiscuous iron-haem-containing enzymesinvolved in oxidative metabolismof a broad variety of xenobioticsand drug compounds.

PHARMACOPHORE

The spatial orientation ofvarious functional groups orfeatures necessary for activity at a biomolecular target.

MULTI-DIMENSIONAL

OPTIMIZATION

The process of paralleloptimization of several relevantdrug-property parameters inconcert with activity, to producea drug candidate with balancedproperty profiles suitable forclinical development.

NUCLEAR MAGNETIC

RESONANCE

A spectroscopy tool used for theassignment and confirmation ofchemical structure of acompound or biologicalmacromolecule. Sophisticatedmulti-dimensional methods areused to characterize larger andmore complex biomolecules.

COMBINATORIAL CHEMISTRY

Synthesis technologies togenerate compound librariesrather than single products.Robotic instruments for solid- and solution-phasechemistry, as well as high-throughput purificationequipment, are applied.

DRUG-LIKENESS

A scoring metric (computational)for the similarity of a givenstructure to a representativereference set of marketed drugs.

Hits/leads

GP

R-1

GP

R-2

GP

R-3

GP

R-4

GP

R-5

GP

R-6

GP

R-8

GP

R-7

GP

R-9

ChemogenomicsNatural products

Random screening

Combinatorial chemistry Priviledged motifs Literature and patents

Ligand design

Endogenous ligand

H2N– –COOHD RV

Y IH

P F

NR'R

N

L

L

+

L

+

Figure 4 | Hit-identification strategies. The most commonly applied hit-identification strategies today range from knowledge-based approaches, which use literature- and patent-derived molecular entities, endogenous ligands or biostructural information,to the purely serendipity-based ‘brute-force’ methods such as combinatorial chemistry and high-throughput screening. Theamalgamation of both extremes is anticipated to deliver more high-content chemical leads in a shorter period of time.



R E V I E W S

design and combinatorial chemistry far more challeng-ing, as synthesis protocols for compound arrays are oftenthe limiting factor in the choice of useful designs. Theclose collaboration of computational scientists andchemists is therefore essential for formulating library pro-posals that fit with the target structure requirements andthat are simultaneously amenable to parallel syntheticassembly. Finally, an understanding of the mechanismof action of a biological target, which is often availablefor many families of enzymes, is an important aid inbiasing compound collections. These mechanism-basedlibraries have been applied successfully to a variety ofproteins to generate transition-state mimics using eitherparallel solution- or solid-phase synthesis techniques23.

Privileged structures or motifs. Another widely usedapproach concerning the generation of targeted com-pound collections is ligand motif-based library design.This is particularly relevant for targets for which verylimited or no biostructural information is available. It ishere that elements of known biologically active mole-cules are used as the core for generating librariesencompassing these ‘PRIVILEGED STRUCTURES’24. Especiallyin the area of GPCRs, such design tactics have beenapplied successfully25. An inherent issue linked to thisapproach is the fact that these motifs can show promis-cuous activity for whole target families, so selectivityconsiderations have to be addressed very early on. Therestricted availability of privileged structures, andresulting issues concerning intellectual property, clearlylimits the scope of this ligand-based approach to someextent. As a result, there is a continued need to identifynovel proprietary chemotypes, and computational toolssuch as Skelgen26 and TOPAS27 have already shown theirpotential in this area.

‘Cherry picking’ from virtual space. A highly sophisti-cated way to avoid the synthesis of trivial analogues isthe application of virtual screening tools in order tosearch through chemical space for topologically similarentities using known actives (seed structures) as refer-ences. In addition, biostructural information can also beapplied if available28. Principally, one can subdivide sucha virtual screening exercise into three main categories,namely virtual filtering, virtual profiling and virtualscreening (BOX 2). The first focuses on criteria that arebased on very fundamental issues concerning pharma-cological targets in general. In this filtering step, all can-didates are eliminated that do not fulfill certain gener-ally defined requirements. These elimination criteriacan either be based on statistically validated exclusionrules, substructural features or on training sets ofknown compounds. A retrospective analysis of drugmolecules that demonstrated appropriate bioavailabilityformed the basis for the ‘rule of five’ guidelines, whichmake use of simple descriptors, such as molecular mass,calculated lipophilicity and hydrogen-bond donors/acceptors, in order to assess the probability of com-pounds being absorbed intestinally29. DEREK and TOP-KAT are prediction tools for toxicological liabilities basedon substructural analysis30. Artificial neural networks

metric, otherwise it is meaningless17. In any case, toincrease the quality of a compound inventory, certainfiltering techniques have to be applied for weeding outcompounds that contain unfavourable chemical motifs.Database-searching tools have been developed thatallow the differentiation of desired and undesired com-pounds. These computational algorithms are oftenbased on sub-structural analysis methods, similarity-searching techniques or artificial neural networks18.Besides the application of those for filtering physicallyavailable compound collections or vendor databases,such algorithms can of course also be used to screen andvalidate virtual combinatorial libraries. It is in this set-ting that computational screening can have the greatestimpact, owing to the overwhelmingly large number ofcompounds that are synthetically amenable usingcombinatorial chemistry technologies (BOX 1).

Focusing for librariesThe ‘combinatorial explosion’ — meaning the virtuallyinfinite number of compounds that are syntheticallytractable — has fascinated and challenged chemists eversince the inception of the concept. Independent of thelibrary designs, the question of which compoundsshould be made from the huge pool of possibilities alwaysemerges immediately, once the chemistry is establishedand the relevant building blocks are identified.

The original concept of ‘synthesize and test’, withoutconsidering the targets being screened, was frequentlyquestioned by the medicinal chemistry community andis nowadays considered to be of much lower interest dueto the unsatisfactory hit rates obtained so far. The daysin which compounds were generated just for filling upthe companies inventories, without taking any design orfiltering criteria into account, have passed. In fact, mostof the early combinatorial chemistry libraries have nowbeen largely eliminated from the standard screening setsdue to the disappointing results obtained after biologi-cal testing. The first generation of such combinatoriallibraries were unattractive for most screening groupsdue to overloaded molecular complexity19, poor drug-like features and low product purity. As a consequence,there is now a clear trend to move away from huge anddiverse ‘random’ combinatorial libraries towardssmaller and focused drug-like subsets. Although thediscussion of how focused or biased a library should beis still an ongoing debate, the low hit rate of large, ran-dom combinatorial libraries, as well as the steadyincrease in demand for screening capacity, has set thestage for efforts towards small and focused compoundcollections instead.

Guided by the target. Biostructural information derivedfrom mutagenesis data, as well as NMR or X-raycrystallographic analysis, has long been used for drugdiscovery purposes. Although the emphasis was initiallyfocused more on single compound synthesis, a shifttowards designing specific compound libraries is morecommonplace today20–22. Recognizing that PARALLEL SYN-

THESIS procedures cannot be applied to every structuralmotif makes the integration of biostructure-based

DIVERSITY

A property–distance metricreflecting the dissimilarity ofobjects (molecules).Variousmolecular descriptors (indices)are used to define compounds ina numerical fashion so that theycan be readily compared. Suchmeasures must be consideredwithin an appropriate context tobe meaningful.

‘FREQUENT-HITTER’ LIABILITIES

An empirically derived metric bywhich compounds are assigned a probability to produce (false)positive results (hits) frequentlyin diverse screening assays.

MOLECULAR TOPOLOGY

A graph-based method ofdescribing molecular structureusing atom connectivity throughthe molecular framework andassigning atoms or substructuraldomains with various propertytypes: lipophilic, H-bondacceptor/donor, positively/negatively charged and so on.

DESCRIPTORS

Metrics used to numericallydescribe a structure or certainmolecular attributes of acompound (for example,Tanimoto, Ghose and Crippen,BCUT and so on).

PARALLEL SYNTHESIS

The process by which a set ofindividual compounds is madesimultaneously using commonchemical building blocks andhomologous reagents.

PRIVILEGED STRUCTURE

A specific core or scaffoldingstructure that imparts a genericactivity towards a protein familyor limited set of its membersindependently of the specificsubstituents attached to it.



R E V I E W S

level of sophistication can be achieved by using three-dimensional pharmacophores; however, this requiresmuch more knowledge in terms of ligand informationand conformation, as well as far greater computa-tional power and time. Generating three-dimensionalCONFORMERS of all library members clearly limits the sizeof the virtual library to a great extent, but having a two-dimensional searching step integrated for pre-selectionhelps to reduce the number of possible candidates.Virtual pharmacophore searches have been widely usedduring recent decades, and the impact within structure-or property-based drug discovery and lead design isbecoming more prevalent35.

The ultimate step of a virtual screening campaignis the introduction of the ‘fourth dimension’, namely,the target structure itself. The closest virtual approachto real bio-screening is virtual DOCKING AND SCORING, inwhich compounds are selected by defining interactionpatterns of virtual compounds with the binding site of

have been described that can discriminate betweendrug-like and non-drug-like compounds31, moleculeswith high likeliness for cytochrome P450 interactions32

or compounds that might show hERG liabilities33 tofurther profile compounds in more depth.

Although the filtering and profiling steps can beapplied to qualify particular compounds as being moreor less drug-like, the virtual screening part insteadencompasses specific project information to predict acertain binding propensity. Computational tools basedon two-dimensional topological descriptors haveproved to be very valuable for rapidly screening hugedatabases using known molecules as seed structures togenerate activity-enriched libraries34. A big advantage inthis context is the fact that besides the speed, a singlemolecule can be sufficient to identify compounds thatare very different structurally, but which show similarbiological activity, out of a particular virtual (or physi-cally available) compound library. Searching at a higher

CONFORMERS

Distinct three-dimensionalforms of a molecular structure of a given atomic connectivity,which results from internalrotations about single bondsbetween atoms.

Box 1 | The issue of chemical space

The number of synthetically tractable compounds can be taken to be practically unlimited. The resulting chemicalspace is hard to comprehend, but the issues encountered are easily exemplified. Benzimidazoles, for example, are oneof many interesting classes of molecules for which chemists can immediately devise various synthesis access routes.The shown example starts from the corresponding modified phenylenediamines by elaboration with carboxylic acids,alkyl- or (hetero)arylmethyl-halides, primary or secondary amines and boronic acids. Having only 100 entities of eachbuilding block available, a library of 100 × 100 × 100 × 100 = 108 benzimidazoles is conceivable. Even though only afraction of those molecules are probably pharmacologically relevant, the huge number of possibilities indicates thatthe compound collection could span a large portion of chemical property space containing members with biologicalactivities against many different pharmacological targets. Obviously the synthesis and testing of all combinations isneither feasible nor meaningful. Virtual screening technologies help to filter out the unfavourable combinations andpredict actives out of such a library proposal if particular target and/or ligand information is available.

R1

R4

N

NR2

R3

O

NHNH NH

BOHHO

B

N

OHHOB

OHHO

O

N

BrBr

Br

O

HO

OO

HO

O

HO

O

N

N

N

N

N

HN

O

O

N

N

N

O

O

N

N

N

N

Virtual filtering,profiling andscreening

Library proposal Virtual selection Library synthesis

Virtualstructures

Proposedstructures

Libr

ary

enum

erat

ion P

arallel synthesis



R E V I E W S

So, the tight integration of database generation(library proposals), virtual screening, synthesis andmultidimensional testing (affinity, selectivity, physico-chemical properties and so on) is mandatory to ensurea successful process. Rather than synthesizing largecompound libraries, only to find out later that hits door do not materialize, rapid feedback loops are farmore useful, as they allow the flexible adaptation ofeither the computational tools, the information usedas input or the library proposals themselves. Adaptivecycles need to be established to guide the ‘journeythrough chemical space’ by computational scientistsand the resulting data generated by chemists and biol-ogists. Needless to say, it is imperative that logisticalhurdles are overcome and that artificial boundariesbetween many disciplines are eliminated for maximizingthe potential success of this approach.

the target protein. The requirement of crystallo-graphic data, detailed knowledge of the binding modeand inherent issues concerning affinity-scoring func-tions still limits this approach to a great extent whenconsidering the screening of large virtual libraries36.The iterative sequence of selection and refinementusing one-dimensional descriptor, two-dimensionalligand and three-dimensional pharmacophore screen-ing reduces the selected candidates to a manageablenumber. This leaves the highest-ranking molecules forfurther filtering by biostructure-based docking andscoring, and so provides both high activity enrichmentand structural novelty.

As promising and valid as many of these computa-tional algorithms are, they clearly can only be regardedas prediction tools. A continuous validation of pro-posed actives by rapid synthesis and testing is essential.

DOCKING AND SCORING

The process of computationallyplacing a virtual molecularstructure into a binding site of abiological macromolecule(docking) and flexibly or rigidlyrelaxing the respective structuresthen ranking (scoring) thecomplementarity of fit.

Box 2 | Virtual screening

Individual steps in a virtual screening cascade can be subdivided into four principal components that distinguish thelevel of complexity delivered as input. First, general criteria are applied to eliminate all chemical structures that possessreliably predicted detrimental features, which would make them intrinsically less attractive as potential drugs.Molecular size, lipophilicity or potential metabolic liabilities, for example, could be used to reduce the number ofpossible candidates significantly. These filters are regarded as one-dimensional, as they are typically scalar and nodetailed information on project-specific criteria is used. Topological searches from known ligands are often successfullyapplied in virtual screening when seeking compounds showing similar biological activities but with different structuralcharacteristics (template hopping). Three-dimensional pharmacophore models are also applied for virtual screeningwhen more detailed information concerning ligands and three-dimensional pharmacophore orientations is available.Even more structural knowledge is required when applying docking to a target structure and scoring the fittest chemicalentities to be prioritized.

Tools

MACCS40

BCUT41

Cats34

Topas27

CATALYST42

Skelgen26

Pro-select43

Ludi44

Input Filter Library

One-dimensional 10n molecules

n molecules

Three-dimensional

Two-dimensional

Four-dimensional

N

N

NN

H2N

F

Targetspecificity

300

250

200

1 2 3 4 5

150

100

50

0

Com

poun

ds

Score



R E V I E W S

in a highly parallel fashion. So, computational tech-nologies play a central role in chemogenomics not onlyfor the generation of biased libraries, but also for theidentification and clustering of biological targets.

The main question still to be answered is whether itis possible to overlay a certain chemistry or topologyspace (defined by the compound libraries) with a partic-ular target space (defined by the sequence of the pro-teins). This depends on reliable biological and chemicalannotation systems, as well as the possibility of linkingthem to each other. Experimental evidence for thechemogenomics concept has been delivered by medicinalchemistry for a long time. It is well known that more orless conservative changes in a molecular structure notonly affects the activity, but also the selectivity, of a com-pound. The same phenomena are observed on the targetside where protein mutations might lead to a completeloss of ligand activity or show no effect at all. In otherwords, the probability and extent of ligand binding canbe tuned by the similarity of the chemical entities on theone hand, as well as by phylogenetic distance of thetargets on the other. Understanding both, and being ableto systematically annotate target- and ligand-space on apharmacologically relevant basis, makes possible theidentification of novel ligands and targets simultane-ously (BOX 3). Receptor de-orphanization is no longerrestricted to the identification of endogenous ligands,but can be achieved in a prospective manner by apply-ing the similarity principle on both the ligand and thetarget side. Once a target protein is identified and corre-lated to a particular family cluster, the testing of focusedcompound libraries biased towards that subfamilyshould deliver hits for this particular target as well. So,the first step within a chemogenomics endeavour is thehunting for novel genes and proteins.

The constant growth in the amount of data emerg-ing from genomics and proteomics studies clearlyrequires alternative methods to support the classicalstrategies for target assessment. Computational methodsare making an increasing contribution, and bioinfor-matics plays a crucial role37. Successful applications ofin silico target identification through bioinformaticapproaches have been described recently in whichsequence-similarity searching was performed usingknown DNA or protein sequences as seed information.To take a specific example, a recent publicationdescribes the identification of four GPCRs fromgenome databases that were searched using variousGPCR sequences as queries. Transcripts for all fourgenes were experimentally detected in the brain, whichindicated that these receptors might be novel targetsfor central nervous system research38. Eventually,focused compound libraries using either the privi-leged structure approach, or more advanced virtualscreening-based compound arrays, should allow us toidentify small-molecule agonists, further validate thetarget and simultaneously move forward with drug-like compounds into the lead-generation and lead-optimization phases. The ongoing debate as towhether two-dimensional versus three-dimensionalor ligand- versus target-based input is required for

ChemogenomicsThere is no doubt that computational tools (both ligand-and biostructure-based tools) can be very useful toprioritize compounds that are more likely to be active ata particular target compared with others. The synthesisof these predicted actives ranked by any similaritymetric results in a set of focused compounds that covera certain portion of chemistry space. Owing to the tar-get-related input that was used for biasing this set ofcompounds, a link to the corresponding proteins isestablished. The fuzziness inherently incorporated dueto the imprecise nature of any prediction method is infact of great benefit for a chemogenomics approach inwhich focused libraries of small drug-like molecules areused for the identification and validation of novel targets

Box 3 | Similarity searching

Similarity-searching algorithms applied in chemo- and bioinformatics serve to identifyand annotate DNA or protein targets, as well as potential small-molecule modulators,at different levels of sophistication. On the basis of genomic information, proteins can be translated from their DNA sequence. Similarly, in chemistry, SMILES or BIT STRINGS areapplied to annotate compounds and, more importantly, large compound databases (thechemical genome of a library). Those can be further classified, for example, by theirthree-dimensional pharmacophore representation, just as proteins can be classified bytheir function. Homology alignments of DNA or proteins by sequence similarity makethe grouping of targets into target families possible. This is analogous to similarity-basedvirtual screening in which compounds are grouped on the basis of their annotation.Matching both topology and target space allows the identification of novel targets andligands simultaneously.

Sequence space

5′–CGU GUC GGC–3′ 1100 1100 1110 1001Annotate

Translate

Classify

Group

Topology space

H2NNH

NOH

O

O

O

NH

NH2

HN

Cl

Cl ON

N

N

mgr_1 rat I A P L Q R V G

mgr_2 rat G P T F R R L G A

F

G

S

mgr_3 rat G P T F R R L G

mgr_4 rat L P T M R R L G

mgr_5 rat A A T L Q R I P

S

SMILES

A character-based line notationfor chemical structures.

BIT STRINGS

A contiguous set of charactersthat consists entirely of 1s and0s, which can be used to encode,for example, the presence or not of structural elements in a compound.



R E V I E W S

full exploitation of high-quality and high-throughputtechnologies in chemistry, biology, molecular propertyanalytics and ADME assessment in an integratedfashion. Continuous efforts are necessary to developeven more reliable, predictive virtual screening toolsand knowledge-based algorithms for a better estimationof the key biophysical and even in vivo toxicological lia-bilities before synthesis is initiated. The combination ofthese tools into an integrated process (FIG. 5) will allowmodern drug discovery to progress to a new level ofsophistication. A quantitative improvement of thesuccess rates is essential to cope with the ever-increasingexpectations on pharmaceutical research, as well as themany new therapeutic targets expected to derive fromintensified genomics and proteomics programmes.

Outlook aheadIn a manner similar to the tremendous developmentand maturation of oligonucleotide chemistry since the1980s to the situation today in which DNA primerscan be ordered by e-mail and are delivered the nextday, we expect that parallel organic chemistry willprogress analogously. This will allow the synthesis offocused compound libraries very rapidly on demand.The enormous chemical space that can already becovered by well-established chemical procedures(being much larger than any compound inventory willever be), linked with ever better virtual screening andprediction tools, will give the chemist the opportunityto propose certain chemotypes to be ‘squeezed’ intothe relevant pharmacology space by appropriate deco-ration. Increasingly, examples will follow where chem-istry is dictated by the chemist and not by chemicalsidentified after random screening. Downstream opti-mization work will become increasingly effective notonly because the chemistry is established, but alsobecause a broader choice of templates and buildingblocks will be readily available. The application ofHTS technologies will certainly move from pure ran-dom testing of huge compound pools to iterative RAPID

FEEDBACK SCREENING of smaller, but more focused, com-pound ensembles. Therefore, the time spent onobtaining the relevant information, rather than thesheer capacity of synthesis and testing, will determinethe success of a research programme. Stepping backfrom the rather ‘safe’ HTS paradigm to discovery invirtual space, which is still not yet fully developed,certainly needs courageous management decisions,not only in terms of financial investments, but also inorganizational evolution. Breaking down artificialboundaries between different disciplines is a prerequi-site for making full use of their potential. The largenumber of emerging targets, which are expected fromfunctional genomics, demands novel and effectiveapproaches for the hit and lead generation process aswell as the lead optimization phase. Therefore chem-istry will increasingly be applied upstream for bothtarget identification as well as assessment39 where therefined tools for investigating receptor pharmacolo-gies will already encompass the properties of futurepotential drugs.

library focusing clearly shows that the computationalcommunity is still in the process of identifying themost valuable tools and strategies at each stage.

ConclusionHit and lead generation are key processes involved inthe creation of successful new medicinal entities, andit is the quality of information content impartedthrough their exploration and refinement that largelydetermines their fate in the later stages of clinicaldevelopment. It is in the early phases of drug discoverythat changes in process, such as the early interceptionof key ADME parameters, can have the maximumimpact on later-stage success and timelines. The presenthigh attrition rates, especially after lead-optimizationphases, indicate that drug discovery as a sequentialalignment of independent disciplines is ineffective fordelivering high-quality medicines of the future, andthat issues beyond activity and selectivity must beaddressed as early as possible in a flexible, parallelfashion. In our view, the combination of virtualscreening and parallel medicinal chemistry, in con-junction with multi-dimensional compound-propertyoptimization, will generate a much-improved basisfor proper and timely decisions about which leadseries to pursue further.

Applying this strategy shifts the bottleneck from hitidentification to lead optimization. Therefore novelprocesses will have to be developed downstream for the

RAPID FEEDBACK SCREENING

Rapid feedback provided byassaying small compound sets(< 1,000) through a medium-throughput assay to guide theSAR for rapid iterative designand synthesis cycles.

Figure 5 | Where there’s a will, there’s a way… The discoveryand development of new medicines is regarded as one of themost complex areas of research in both industry and academia.The expertise of many disciplines is essential to resolve themultifaceted challenges facing this discovery endeavour, rangingfrom pathway analysis to late-stage clinical development. In anincreasingly complex and fast-paced research environment, thetight integration of complementary disciplines and technologiesis becoming more essential than ever. This process will expandour current understanding of drug discovery, necessitating amove away from serial programmes towards increasingly multi-parametric parallel processing.



R E V I E W S

1. Drews, J. Drug discovery: A historical perspective. Science287, 1960–1964 (2000).

2. Lander, E. S. et al. Initial sequencing and analysis of thehuman genome. Nature 409, 860–921 (2001).

3. Knowles, J. & Gromo, G. Target selection in drug discovery.Nature Rev. Drug Discov. 2, 63–69 (2003).

4. Hopkins, A. L. & Groom, C. R. The druggable genome.Nature Rev. Drug Discov. 1, 727–730 (2002).

5. Lenz, G. R., Nash, H. M. & Jindal, A. Chemical ligands,genomics and drug discovery. Drug Discov. Today 5,145–156 (2000).

6. Hodgson, J. ADMET — turning chemicals into drugs.Nature Biotechnol. 19, 722–726 (2001).

7. Proudfoot, J. R. Drugs, leads, and drug-likeness: An analysisof some recently launched drugs. Bioorg. Med. Chem. Lett.12, 1647–1650 (2002).

8. Alanine, A., Nettekoven, M., Roberts, E. & Thomas, A. Leadgeneration — enhancing the success of drug discovery byinvesting into the hit to lead process. Combin. Chem. HighThroughput Screen. 6, 51–66 (2003).

9. Boguslavsky, J. Minimizing risk in ‘Hits to Leads‘. DrugDiscov. & Develop. 4, 26–30 (2001).

10. Bleicher, K. H. Chemogenomics: bridging a drug discoverygap. Curr. Med. Chem. 9, 2077–2084 (2002).

11. Bajorath, J. Integration of virtual and high-throughputscreening. Nature Rev. Drug Discov. 1, 882–894 (2002).This review article covers the current concepts ofintegrating both virtual and high-throughput screening.

12. Teague, J. S., Davis, A. M., Leeson, P. D. & Oprea, T. Thedesign of leadlike combinatorial libraries. Angew. Chem. Int.Ed. Engl. 38, 3743–3748 (1999).

13. Walters, P. & Murcko, M. A. Prediction of ‘drug-likeness’Adv. Drug Deliv. Rev. 54, 255–271 (2002).

14. Martin, E. J. & Critchlow, R. E. Beyond mere diversity:tailoring combinatorial libraries for drug discovery. J. Comb.Chem. 1, 32–45 (1999).

15. Menard, P. R., Mason, J. S., Morize I. & Bauerschmidt, S.Chemistry space metrics in diversity analysis, library designand compound selection. J. Chem. Inf. Comput. Sci. 38,1204–1213 (1998).

16. Roche, O. et al. Development of a virtual screening methodfor identification of ‘Frequent Hitters’ in compound libraries.J. Med. Chem. 45, 137–142 (2002).

17. Balkenhohl, F., von dem Busche-Hünnefeld, C., Lansky, A.& Zechel, C. Combinatorial synthesis of small organicmolecules. Angew. Chem. Int. Ed. Engl. 35, 2288–2337(1996).

18. Böhm, H.-J. & Schneider, G. (eds). Virtual Screening forBioactive Molecules (Wiley–VCH, Weinheim, 2000).An excellent compendium of current virtualscreening methods.

19. Hann, M. M., Leach, A. R. & Harper, G. Molecular complexityand its impact on the probability of finding leads for drugdiscovery. J. Chem. Inf. Comput. Sci. 41, 856–864 (2001).

20. Crossley, R. From hits to leads, focusing the eyes ofmedicinal chemistry. Modern Drug Discov. 5, 18–22 (2002).

21. Van Dogen, M., Weigelt, J., Uppenberg, J., Schultz, J. &Wikström, M. Structure-based screening and design in drugdiscovery. Drug Discov. Today 7, 471–477 (2002).

22. Carr, R. & Jhoti, H. Structure-based screening of low affinitycompounds. Drug Discov. Today 7, 522–527 (2002).

23. Huang, L., Lee, A. & Ellman, J. A. Identification of potent andselective mechanism-based inhibitors of the cysteineprotease cruzain using solid-phase parallel synthesis. J. Med.Chem. 45, 676–684 (2002).

24. Patchett, A. A. & Nargund, R. P. Privileged structures — anupdate. Annu. Rep. Med. Chem. 35, 289–298 (2000).

25. Bleicher, K. H., Wütherich, Y., Adam, G., Hoffmann, T. &Sleight, A. J. Parallel solution- and solid-phase synthesis ofspiropyrrolo-pyrroles as novel NK-1 receptor ligands.Bioorg. Med. Chem. Lett. 12, 3073–3076 (2002).

26. Stahl, M. et al. A validation study on the practical use ofautomated de novo design. J. Comput.-Aided Mol. Des. 16,459–478 (2002).

27. Schneider, G. et al. Virtual screening for bioactive moleculesby de novo design. Angew. Chem Int. Ed. Engl. 39,4130–4133 (2000).

28. Schneider, G. & Böhm, H.-J. Virtual screening and fastautomated docking methods. Drug Discov. Today 7, 64–70(2002).

29. Lipinski, C., Lombardo, F., Dominy, B. & Feeney, P.Experimental and computational approaches to estimatesolubility and permeability in drug discovery anddevelopment settings. Adv. Drug Deliv. Rev. 23, 3–25 (1997).A landmark publication based on retrospective dataanalysis for bioavailability resulting in the ‘rule-of-five’.

30. Cariello, N. F. et al. Comparison of the computer programsDEREK and TOPKAT to predict bacterial mutagenicity.Mutagenesis 17, 321–329 (2002).

31. Sadowski, J. & Kubinyi, H. A scoring scheme fordiscriminating between drugs and nondrugs. J. Med.Chem. 41, 3325–3329 (1998).

32. Zuegge, J. et al. A fast virtual screening filter for cytochromeP450 3A4 inhibition liability of compound libraries. Quant.Struct.-Act. Relat. 21, 249–256 (2002).

33. Roche, O. et al. A virtual screening method for prediction ofthe hERG potassium channel liability of compound libraries.Chembiochem 3, 455–459 (2002).

34. Schneider, G., Neidhart, W., Giller, T. & Schmid, S. ‘Scaffoldhopping’ by topological pharmacophore search: acontribution to virtual screening. Angew. Chem Int. Ed. Engl.38, 2894–2896 (1999).

35. Mason, J. S., Good, A. C. & Martin, E. J. 3-DPharmacophores in drug discovery. Curr. Pharm. Des. 7,567–597 (2001).

36. Bissantz, C., Folkers, G. & Rognan, D. Protein-based virtualscreening of chemical databases. 1. Evaluation of differentdocking/scoring combinations. J. Med. Chem. 43,4759–4767 (2000).

37. Duckworth, D. M. & Sanseau, P. In silico identification ofnovel therapeutic targets. Drug Discov. Today 7, 64–69(2002).

38. Lee, D. K. et al. Identification of four human G-protein-coupled receptors expressed in the brain. Mol. Brain Res.86, 13–22 (2001).This paper describes the successful identification oforphan G-protein-coupled receptors initiated bybioinformatic approaches.

39. Alaimo, P. J., Shogren-Knaak, M. A. & Shokat, K. M.Chemical genetic approaches for the elucidation ofsignaling pathways. Curr. Opin. Chem. Biol. 5, 360–367(2001).

40. McGregor, M. J. & Pallai, P. V. Clustering of largedatabases of compounds: using the MDL “keys” asstructural descriptors. J. Chem. Inf. Comp. Sci. 37,443–448 (1997).

41. Stanton, D. T. Evaluation and use of BCUT descriptors inQSAR and QSPR studies. J. Chem. Inf. Com. Sci. 39,11–20 (1999).

42. Sprague, P. W. Automated chemical hypothesis generationand database searching with CATALYST. Perspect. DrugDiscov. Design 3, 1–20 (1995).

43. Liebeschuetz, J. W. et al. PRO_SELECT: combiningstructure-based drug design and array-based chemistry forrapid lead discovery. 2. The development of a series ofhighly potent and selective Factor Xa inhibitors. J. Med.Chem. 45, 1221–1232 (2002).

44. Boehm, H.-J. Prediction of binding constants of proteinligands: a fast method for the prioritization of hits obtainedfrom de novo design or 3D database search programs. J. Comput.-Aided Mol. Des. 12, 309–323 (1998).

AcknowledgementDr. Simona Ceccarelli is cordially thanked for providing the cartoons‘Don’t panic….’ and ‘Where there’s a will, there’s a way…’.

Online links

FURTHER INFORMATIONSociety for Biomolecular Screening:http://www.sbsonline.comAccess to this interactive links box is free online.

A guide to drug discoveryHit and lead generation: beyond high-throughput screening

Documents