Vol 437|22 September 2005

NEWS & VIEWS FEATURE

491

PHARMACEUTICALS

A new grammar for drug discoveryMark C. Fishman and Jeffery A. Porter

To realize the potential of the genome for identifying candidate drugs we must move beyond individualgenes and proteins. The signalling pathways in cells provide the right level for such analyses.

Drug discovery is a tough business scientifi-cally, the traditional steps taking nearly adecade. In this article we are concerned onlywith the very beginning of the process, theidentification of a target for a new drug. It isthis step, the choice of a gene product of clini-cal relevance, that is the greatest impedimentto expanding the pharmaceutical arsenal. Inthe United States, only 20–30 new chemicalentities are approved as drugs each year, andthe picture is much the same in Europe. Ofthese, only a quarter act on targets not alreadyhit by an existing drug. Why has the decipher-ing of the 25,000 or so genes in the humangenome not swelled the ranks of new targets?

The major reason, of course, is that targetsare of value for drug discovery only if they canbe convincingly related to disease. Such valida-tion really takes not the one year shown onstandard pipeline charts (Fig. 1), but decades.For example, the discovery of statins as agentsthat lower cholesterol levels in the bloodstreamrested on investigations that began 40 yearsearlier with a study showing the relationshipbetween cholesterol level and vascular disease1.The recent discovery of Gleevec, which hasrevolutionized the therapy of chronic myeloidleukaemia, certain gastrointestinal tumoursand other cancers, was based on work thatbegan in the 1950s (ref. 2). Thus, it has beenargued that well-validated targets, the low-hanging fruit as it were, are simply exhausted3,4.

Validation is generally not a one-step pro-cess, but is rather an edifice built from studiesin epidemiology and disease physiology, andfrom the results of research with animal mod-els. The one exception is mendelian disease. In these disorders, such as cystic fibrosis andsickle-cell anaemia, the inheritance of a muta-tion in a single gene can be incontrovertiblylinked to a physical characteristic, or phenotype— in this case the disease. Of the 6,000 or so illnesses with a mendelian pattern of inheri-tance, the gene responsible has been identifiedin approximately 1,200 (refs 5, 6). Historically,pharmaceutical companies have not concen-trated on these diseases, in some cases becausethe affected protein is not tractable to pharma-ceutical approaches, in others, perhaps, becausethe number of people affected is small. But the

powerful role of a single gene in mendelian dis-ease can provide insight into complex diseaseswhere the same gene accounts for part of thephenotype. Statin therapy, for example, was ini-tially directed to patients with a genetic predis-position to excessive levels of blood cholesterol.But after the drug’s efficacy and safety had beentested, the treatment was extended to a widerpopulation of patients who had the same con-dition but due to many causes1,7,8.

In seeking new targets, we first need to thinkabout how we describe disease. Medical text-books are organized by organ system, and diseases are classified by their pathology orphysiology. If we hope to come to grips withthe heterogeneity of common disease and itsconsequences for the choice of targets for drugdiscovery, clinical manifestations of diseasemust be causally related to a molecular definition. A simple concatenation of all themolecular changes in a disease, for example a compilation of profiles of which genes are being transcribed, and when and where,would be chaotic. A further step of integrationand interpretation is needed. Ideally, thiswould provide a ‘grammar’ that would applyright through from the discovery of a drug tar-get, to its testing in animal models and finallyto the treatment of patients. In this context, agrammar means a set of rules, not for organiz-ing words into sentences but for translatinggene products into medicines. Just as the

grammar we use in our speech is embedded inlanguage, a grammar of drug discovery shouldbe embedded in biological systems.

Molecular pathways Intracellular molecular signalling pathwaysprovide such a grammar for the genome.These pathways are triggered by extracellularmolecules that bind to receptors in the cellmembrane, thereby switching on relay systemsinside the cell. The upshot is gene activation,or inactivation, that affects a cell’s behaviour —its ability to grow or differentiate, for example,or to undergo division or self-destruction. Thenumber of distinct signalling pathways recog-nized in human cells depends on the defini-tion. If classified by the type of cell-surfacereceptor involved, it could be as low as 16 (ref.9). Using definitions based on cell type orgene-family member, and variants thereof, itcould be as many as 200 (ref. 10). In any case,the number of signalling pathways is far fewerthan the 25,000 or so human genes.

The elements of four now-canonical molec-ular pathways are shown in Figure 2 (overleaf):these lead from the receptors for proteins ofthe Wnt, Hedgehog (Hh), transforming growthfactor-� (TGF-�) and insulin/insulin-likegrowth factor (IGF) families. Only the corecomponents are depicted here in a much sim-plified manner. The important point is thatthese pathways are conserved throughout

Target discovery/validation

Assay development

High-throughput screenLead selection

Lead optimization/candidate selection

Preclinical testing

Clinical trials

10 2 3 4 5 6 7 8

Years

Figure 1 | Drug-discovery timeline. The initial phase, where promising targets are first discovered and linked to a disease, is shown here as taking a year, but it can require several years of work in bothuniversity and industrial labs. Subsequent stages involve screening for and identifying compounds that safely alter the activity of a particular target, before their efficacy is tested in clinical trials.

22.9 N&V Feat Fishman MH 16/9/05 5:51 PM Page 491

Nature Publishing Group© 2005

NEWS & VIEWS FEATURE NATURE|Vol 437|22 September 2005

492

most of the animal kingdom, in invertebratesand vertebrates, and so they evidently controlsome of the basic cellular functions of life.

The power of molecular pathways to providea grammar is evident from studies of embryonicdevelopment. As first demonstrated in the fruit-fly Drosophila, the language of development is best phrased by the grammar of molecularpathways11. At a molecular level, the character-istic units (sentences, if you will) of fly develop-ment are written in the form of pathways thatdetermine, for example, the first anterior–posterior patterning decision, or the subsequentdivision of the body plan into segments or thegeneration of wings, eyes and other tissues. Inzebrafish and mouse, similar unitary modules,often using the same molecular pathways, alsodefine elements of organ systems and functionin embryonic development12.

Fundamental cellular processes that controlgrowth and cell death in adults are similarlydefined by canonical molecular pathways13,14.Such pathways can be represented graphically10

and measured quantitatively15. The order andinteractions of pathway components can bedefined by genetics and by protein chemistry.All in all, they can provide a grammar appli-cable to drug discovery and medicine, from target validation through to the clinic.

The link between disease and signallingpathway is best validated for genetic disorders.Table 1 lists several diseases caused by muta-tions in the genes for components of the Wnt,Hh, TGF-� and insulin/IGF pathways. Pertur-bation of the essential processes driven bythese pathways is the cause of many diseases,including diseases with complex causes, suchas diabetes and heart disease.

Perhaps the best example known so far is theperturbation of a single pathway that regulatescell growth — the insulin/IGF–AKT pathway(Fig. 2) — in a host of apparently unrelated dis-eases, and how targeting of a key node in thispathway can have beneficial effects in all thosediseases13. The pathway is activated in manycancers. For example, an increase in levels ofphosphatidylinositol 3-kinase (PI3K) or muta-tions in its regulatory subunit cause cancers ofthe ovary and stomach, and activation of AKTor a decrease in PTEN function cause tumoursof the prostate gland.

Other growth phenomena also depend onthe insulin/IGF pathway. Examples are theproliferation of smooth muscle that leads to anarrowing of arteries (restenosis) followingtreatment to unblock them, and the increase inthe numbers of certain immune cells — T cells— during transplant rejection. A key down-stream node in this pathway in all these disorders is the protein mTOR, which regu-lates the translation of messenger RNA into protein. In clinical trials, inhibition of mTORwith derivatives of the macrolide rapamycin,an immunosuppressive drug, is now provingeffective against cancer and restenosis as wellas in transplant rejection.

The key prediction of such logic is that the

drug target can be a ‘sensitive node’, not neces-sarily the protein that is specifically perturbed— which may be hard to define or not readily‘druggable’ using available treatments. Forexample, predisposition to colon cancer is duein some cases to inherited mutations in APC, acomponent of the Wnt pathway (Fig. 2), a pro-tein that is technically difficult to target. Butcompounds that block downstream interac-tions in this pathway, those of �-catenin withits co-activators, have been discovered, and areable to suppress growth of tumours arisingbecause of mutations in APC16,17.

How can we discover the druggable nodes in other pathways? One approach is to selectpoints of intervention, based on the currentunderstanding of the pathways and their drug-gable components, and begin building cell-freeassays to screen for small molecules that willalter the activity of these targets. A less biasedapproach is to combine genetics with chemicalscreens using potential drugs. The premise ofsuch an approach is that the effect of a drugresembles that of the genetic perturbation of itstarget18–20. For example, screens in yeast haveshown that the growth-inhibitory effect of acompound often matches that of a mutation inthe gene encoding its target21,22. In the chemicalscreen, libraries of drug-like compounds areapplied to cells, or even whole organisms, toalter pathway activity, and their effects are thencompared with those of known mutations. The presumptive target can be confirmed

biochemically. Thus, in principle, a compen-dium of drugs could be established that inter-rupt or stimulate every pathway of interest.

Next steps‘In principle’ is the crucial phrase here, for ourcurrent understanding of molecular pathwaysis insufficient as a platform for effective phar-maceutical discovery. Several biotechnologycompanies have focused on the known elements of a few key pathways to target themwith new medicines. But for the genome to betranslated into medicines with any reliabilityand regularity, far more work needs to bedone. Defining the role of pathways in com-plex diseases will undoubtedly take manyyears. A first step is to define the full complex-ity of these signalling networks at a molecularlevel — their ‘systems biology’ — includingactivity specific to a particular cell type,dynamic feedback mechanisms, extensiveinter-pathway connectivity, the kinetics of sig-nalling, and, of course, their state of activationin disease. The following issues will have to beaddressed.

● Pathway outputs have different effects in dif-ferent contexts. For example, the pathway acti-vated by another extracellular signal, fibroblastgrowth factor (FGF), controls cell division inmany tissues but in others it regulates cell differentiation, shape or survival23. Sonichedgehog, one of the related Hh signalling

Figure 2 | Four signalling pathways linked to disease. The diagram depicts the main components of theWnt, Hedgehog (Hh), transforming growth factor-� (TGF-�) and insulin/insulin-like growth factor(IGF) pathways within a single cell. In each case, triggering a receptor on the cell surface activatesmolecular relay systems, which, on transmission of the signal into the nucleus, result in genes beingswitched on or off. The arrows indicate activation and the T-bars show inhibition. Misregulation of several of these components is directly associated with various inherited and sporadic illnesses (see Table 1). For simplicity, details of the component nomenclature and known interactions are not covered here (see ref. 10).

22.9 N&V Feat Fishman MH 16/9/05 5:51 PM Page 492

Nature Publishing Group© 2005

NATURE|Vol 437|22 September 2005 NEWS & VIEWS FEATURE

493

molecules, defines cell fates in the embryonicneural tube, but promotes cell division in theembryonic cerebellum24. The outcome mayeven vary for a single tissue. For example,when it acts in breast epithelium, TGF-�normally helps to prevent the malignant trans-formation of the cells, in that it hinders cellgrowth and multiplication and promotes cellsuicide. But if the epithelial cells are trans-formed as a result of a mutation upstream ofSMAD, for example, TGF-� enhances blood-vessel formation and tumour-cell invasiveness,which encourage tumour growth and disper-sal25. Understanding the druggable elementsor nodes in pathways that can trigger suchaltered biological states should define new targets for treating many different diseases. ● Pathways are dynamic, with feedback mech-anisms. For example, FGF stimulation of air-way formation in Drosophila is terminated by the intracellular activation of the proteinSprouty, which restricts the process to a specific location23,26. Similarly, Wnt signallingcauses the up-regulation of inhibitors of theWnt receptor Fzd, which then inhibit the path-way (Fig. 2)27. In organ systems, changes infeedback, or in its phasing with regard to thestimulus, can have adverse effects. For exam-ple, abnormal breathing patterns result whenfeedback to the chemoreceptor concerned isout of phase with its output. It is likely thatfeedback abnormalities have adverse conse-quences in a cellular context as well. A majorgoal is to define the molecular componentsthat are responsible for maintaining physio-logical balance, as these proteins are likely tobe key regulators of disease.● Pathways intersect. Work in classical genet-ics has defined the phenotypic consequenceswhen different sites on the same pathway areperturbed11. Pathways also engage in extensivecross-talk28,29 with each other, however, andthese interactions change over time, for exam-ple after stimulation. But despite the complex-ity of these interactions, they are finite: theycan be mapped and the key nodes in the inter-sections identified. As we come to understand

this cross-talk and its role in disease, a logic for combination therapy should emerge inwhich two or more pathways can be targetedsimultaneously.● Pathway kinetics vary between cells. Thequantitative flow of signals through pathwaysdepends on the levels of the specific compo-nents of a pathway. For example, quantitativemodelling shows that the level of axin (Fig. 2)is what determines the amplitude and dura-tion of signals through the Wnt pathway15. Tocreate cell-specific modulators of pathways wewill probably need to define these limitingsteps and target them.

Clinical relevanceOne advantage of a pathway-based clinical taxonomy is that it affords a mechanistic foun-dation for disease description. Historically, diseases have been categorized by organpathology. Today we are moving towards pan-genomic assays of gene expression, protein levels and the ways in which proteins becomemodified after they have been produced frommRNA. The results of these assays become easier to interpret when several elements of a pathway are affected. Subtle changes in anindividual protein may be below the thresholdof detection, but may be amplified by sophisti-cated computational methods that incorporateseveral pathway components30.

Another advantage of focusing on pathwaysis the ability to choose the most appropriate setof patients for initial tests of possible treat-ments. For many pathways, people with specificgenetic defects have been described (Table 1),but, because of their rarity, the discovery of atreatment has often been relegated to the‘orphan drug’ category. The development oftherapies for such patients would not only servea medical need, but would often be readilyextrapolated to a wider population. For exam-ple, Gorlin’s syndrome, which results from mu-tations in the protein Patched-1 (Ptc in Fig. 2; a negative regulator of Hh signalling), causes a brain cancer called medulloblastoma and acommon form of skin cancer. The syndrome

itself only affects between 1 in 50,000 and 1 in150,000 people31. But many other tumours,including non-small-cell lung cancer, gut-related tumours and prostate cancer, depend onHh signalling, even though they do not seem tohave mutations in pathway components32.

The rationale for extrapolating from geneticto sporadic illness is that nature is conservative:this is a safe bet. And there would be immenseimmediate benefit in tackling the rare diseasesin themselves. Patients with uncommon disor-ders are often neglected, their only hope beingthat medicines designed and marketed to treatcommon disorders might coincidentally beable to treat theirs. In contrast, in the logic ofmolecular pathways, such patients would beviewed as key intermediaries in the drug-dis-covery process, providing proof of concept.Along the undoubtedly long route to makingmolecular pathways a useful grammar formedicine, essential steps will involve success-fully treating these well-defined but rare dis-eases. This approach could not only bring anew order to the genome, but could also have asalutary ‘side effect’ — a refocusing of drug-discovery research on neglected diseases. ■

Mark C. Fishman and Jeffery A. Porter are at theNovartis Institutes for BioMedical Research, 250 Massachusetts Avenue, Cambridge,Massachusetts 02139, USA. e-mail: mark.fishman@novartis.com

1. Tobert, J. A. Nature Rev. Drug Discov. 2, 517–526 (2003).2. Nowell, P. C. & Hungerford, D. A. Science 132, 1487–1501

(1960).3. Drews, J. Science 287, 1960–1964 (2000).4. Brown, D. & Superti-Furga, G. Drug Discov. Today 8,

1067–1077 (2003).5. Hamosh, A. et al. Nucleic Acids Res. 30, 52–55 (2002).6. Schuener, M. T., Yoon, P. W. & Khuory, M. J. Am. J. Med.

Genet. 125C, 50–65 (2004).7. Mabuchi, H. et al. N. Engl. J. Med. 305, 478–482 (1981).8. Mabuchi, H. et al. N. Engl. J. Med. 308, 609–613 (1983).9. Gerhart, J. & Kirschner, M. W. Cells, Embryos, and Evolution

(Blackwell, Malden, MA, 1997).10. www.biocarta.com/genes/allPathways.asp11. Nüsslein-Volhard, C. & Wieschaus, E. Nature 287, 795–801

(1980). 12. Harris, T. E. & Lawrence, J. C. Jr Science STKE 2003, re15

(2003).13. Metzstein, M. M., Stanfield, G. M. & Horvitz, H. R.

Trends Genet. 10, 410–416 (1998).14. Fishman, M. C. & Olson, E. N. Cell 91, 153–156 (1997).15. Lee, E., Salic, A., Kruger, R., Heinrich, R. & Kirschner, M. W.

PLoS Biol. 1, 116–132 (2003).16. Emami, K. H. et al. Proc. Natl Acad. Sci. USA 101,

12682–12687 (2004).17. Lepourcelet, M. et al. Cancer Cell 5, 91–102 (2004).18. Stockwell, B. R. Nature Rev. Genet. 1, 116–125 (2000).19. Giaever, G. et al. Nature Genet. 21, 278–283 (1999).20. Hughes, T. R. et al. Cell 102, 109–126 (2000).21. Lum, P. Y. et al. Cell 116, 121–137 (2004).22. Parsons, A. B. et al. Nature Biotechnol. 22, 62–69 (2004).23. Tsang, M. & Dawid, I. B. Science STKE 2004, pe17 (2004).24. Ruiz i Altaba, A., Nguyen, V. & Palma, V. Curr. Opin. Genet.

Dev. 13, 513–521 (2003).25. Siegel, P. M. & Massagué, J. Nature Rev. Cancer 3, 807–821

(2003).26. Hacohen, N., Kramer, S., Sutherland, D., Hiromi, Y. &

Krasnow, M. A. Cell 92, 253–263 (1998).27. Logan, C. Y. & Nusse, R. Annu. Rev. Cell Dev. Biol. 20,

781–810 (2004).28. Bouwmeester, T. et al. Nature Cell Biol. 6, 97–105 (2004).29. Giot, L. et al. Science 302, 1727–1736 (2003).30. Mootha, V. K. et al. Nature Genet. 34, 267–273 (2003).31. www.emedicine.com/ped/topic890.htm32. Berman, D. M. et al. Nature 425, 846–851 (2003).

Pathway Associated diseases Affected gene

Hedgehog Gorlin’s syndrome Ptc1

Basal cell carcinoma Ptc1

Medulloblastoma Ptc1

Glioblastoma Gli1

Wnt Vitreoretinopathy Fzd4

Norrie’s disease NDP

Colon APC

Osteoporosis-pseudoglioma/osteopetrosis LRP5

TGF-β Fibrodysplasia ossificans BMP4

Haemorrhagic telangiectasia Alk1

Pulmonary hypertension BMPR2

Juvenile polyposis syndrome SMAD4

Pancreatic cancer SMAD4

Insulin/IGF Cowden’s disease PTEN

Tuberous sclerosis TSC1

Multiple cancers PTEN, PI3K, AKT

Table 1 | Fundamental pathways and a partial list of disease links.

22.9 N&V Feat Fishman MH 16/9/05 5:51 PM Page 493

Nature Publishing Group© 2005