Jump to main content=20
Jump to navigation=20
nature.com homepage
Publications A-Z inde= x
Browse by subject=20

Welcome back: Giovanni Camporeale
Logout of=20 nature.com

=20

Journal = home > Archive >=20 Review= > Full = Text

Review

Nature Reviews Genetics = 8, 601-609 = (August 2007) | = doi:10.1038/nrg2137

Mining gene expression profiles: expression signatures as = cancer=20 phenotypes

Joseph R. Nevins¹=20 & Anil Potti 1 =20 Abo= ut the=20 authors

To= p of page

Abstract

Many examples highlight the power of gene expression = profiles, or=20 signatures, to inform an understanding of biological phenotypes. This is = perhaps=20 best seen in the context of cancer, where expression signatures have = tremendous=20 power to identify new subtypes and to predict clinical outcomes. = Although the=20 ability to interpret the meaning of the individual genes in these = signatures=20 remains a challenge, this does not diminish the power of the signature = to=20 characterize biological states. The use of these signatures as surrogate = phenotypes has been particularly important, linking diverse experimental = systems=20 that dissect the complexity of biological systems with the in=20 vivo setting in a way that was not previously feasible.

View At= a=20 Glance

The capacity to understand biological complexity is = often limited=20 by the ability to define relevant phenotypes. There is perhaps no better = example=20 of this challenge than that seen in cancer. The complexity of the = oncogenic=20 process, involving the somatic acquisition of large numbers of mutations = coupled=20 with variability in the host's genetic constitution, produces a disease = of=20 enormous complexity. Indeed, it is probably not an exaggeration to = suggest that=20 100 breast cancer patients represent 100 distinct diseases. The ability = to=20 dissect this complexity to understand the unique characteristics of the=20 individual patient is key in developing effective therapeutic = strategies.

Traditional methods of characterizing tumours rely on = gross visual=20 information (size of the tumour, degree of spread, histological = characteristics=20 of the tumour) along with a limited number of biochemical assays (for = hormone=20 receptors, growth factor receptors, nuclear antigens, and so on). = Although these=20 tools do provide a way to define tumour subgroups with distinct biology, = it is=20 abundantly clear that these classifications=20 are imprecise, creating heterogeneous groupings of tumours and patients. = Numerous examples show that expression profiles can dissect this = heterogeneity,=20 beginning with work that used expression to distinguish acute myeloid = and acute=20 lymphoblastic leukaemias¹,=20 and then followed by studies that identified previously unrecognized = categories=20 of diffuse large B-cell lymphoma².=20 This capacity to identify otherwise unrecognized biological distinctions = lies in=20 the scale of the analysis =97 a measure of 30,000 or more genes, = reflecting the=20 activity of the entire genome. But the scale goes well beyond that of = 30,000=20 data points given the opportunity to identify patterns of gene = expression,=20 patterns that can be dynamic in response to both physiological and=20 pathophysiological processes.

The advent of technology to measure gene expression on a = genome-wide scale has transformed biology, perhaps more so than the = advent of=20 molecular biology in the 1970s. Genome-scale data has turned biology = into a=20 'data science', now amenable to the power of various forms of = quantitative and=20 statistical analysis that were previously applied to complex data from = finance,=20 meteorology and other fields. Here we focus attention on one particular = aspect=20 of this transformation =97 the ability to develop gene expression = signatures that=20 reflect cancer phenotypes and the capacity to use these signatures as = surrogates=20 for these phenotypes.

The = concept of gene=20 expression signatures

An expression signature is simply a representation of a = biological=20 state, in the form of a pattern of gene expression that is unique to = that=20 circumstance. Underlying the concept is the realization that virtually = any=20 biological condition, whether a developmental state, a cellular response = to=20 extracellular ligands or a pathological state, is reflected in changes = in gene=20 expression. Although no single gene would have the power to define the=20 biological state, the ability to measure and identify patterns of gene=20 expression provides an opportunity to develop these signatures that = reflect=20 biological phenotypes. In a sense, the expression signature becomes a = surrogate=20 representation of the biological phenotype.

The power of the expression signature is twofold. First, = the=20 enormous complexity of the expression data that can be sampled provides = the=20 opportunity to identify patterns of expression that reflect subtle = distinctions=20 in biology. The main limitation lies in the capacity to define the = biological=20 state of interest, whether through the generation of distinct = experimental=20 states that can then be used to train an analysis of expression, or by = taking=20 advantage of existing biological conditions that can be used as the = training=20 opportunity. Second, the expression signature is portable in the sense = that it=20 can be assayed in varied contexts =97 an expression signature that is = developed in=20 a cell-culture context can be measured in a tumour or a histological = section. As=20 such, the expression signature provides a link between otherwise = heterologous=20 systems. A cell-culture phenotype, such as pathway activation, is = difficult to=20 represent in a diverse sample such as a tumour. By contrast, an = expression=20 profile provides a mechanism to link these two states =97 the profile = represents=20 pathway activation in the cell culture that can then be used to = interrogate the=20 expression data from a tumour. In a sense, the gene expression signature = becomes=20 the common currency to connect the experimental state with the in=20 vivo state.

Developing a useful=20 signature

Key to the development of useful expression signatures = is the=20 experimental condition. In many instances, these signatures are = generated in vitro, using experimental settings that can be = carefully=20 controlled to define a specific biological process. To understand the = concept of=20 signatures, consider the cancer cell phenotype of anchorage-independent = growth.=20 This has been used historically as an experimental measure to model the = invasive=20 properties of tumour cells³.=20 To develop an expression signature that is characteristic of this = property, one=20 could select a series of cancer cell lines that varied in their capacity = to grow=20 in soft agar, a measure of anchorage independence (Fig= . 1).=20 RNA is prepared from the various cell lines, both those that exhibit=20 anchorage-independent growth and those that do not, and used for DNA = microarray=20 analysis. These gene expression data are then used in a supervised=20 analysis (see below) to identify a pattern of gene expression =97 a = signature=20 =97 that reflects and, importantly, predicts=20 the phenotype of anchorage-independent growth. The extent to which the = profile=20 actually reflects the biological state, rather than what would be = observed by=20 chance, can be assessed through statistical measures of predictive = ability,=20 typically a leave-one-out cross validation procedure, in which the = profile is=20 generated from a set of samples with one held out and then used to = predict the=20 state of that one sample. Further validation makes use of independent = samples=20 that reflect the phenotype of interest but were not used in the = development of=20 the signature. The ability of the signature that is developed from the = initial=20 training set to then accurately predict the set of test samples that is = used for=20 validation provides an important measure of the signature's = robustness.

= Figure=20 1 | Generation of an expression = signature.

A collection of cancer cell lines are assayed for growth in soft agar = as a=20 measure of anchorage-independent growth. RNA is prepared from the cells = prior to=20 the assay and used for DNA microarray analysis. These data are then used = for a=20 supervised analysis in which a signature is derived that distinguishes = the two=20 cell types (AI+ versus AI-). The extent to which the signature truly = reflects=20 the phenotype is assessed by leave-one-out cross validation to measure = the=20 predictive capacity of the signature.

= High=20 resolution image and legend (81 KB)
= Figures=20 and tables index
D= ownload=20 Power Point slide (279 KB)

Importantly, this signature provides an opportunity to = interrogate=20 other samples, including tumour samples, to assess the extent to which = they=20 exhibit the signature. In this case, one simply uses the tumour sample, = or=20 whatever other set of samples one wishes to profile, as a validation = case,=20 measuring the probability that the microarray data that is derived from = the=20 sample shows the pattern that is defined by the training set that = generated the=20 signature. If done across a set of samples, such as a collection of = tumour=20 samples, one can generate a spectrum of probabilities of the signature. = In the=20 example in Fig= . 2,=20 these probabilities have been depicted as a heatmap, where red = represents high=20 probability and blue represents low probability. Moreover, if the set of = samples=20 is interrogated with multiple signatures, the results can be used for=20 clustering, in much the same way as one would use the primary gene = expression=20 data to cluster samples. In this case, the clusters represent patterns = of=20 signatures across the samples.

= Figure=20 2 | Identification of patterns of=20 signatures.

In this example, a series of five signatures are developed to = represent=20 various biological states. These are then used to predict the status of = the=20 signature in a collection of tumour samples. The predictions are = expressed as a=20 probability measure reflecting the extent to which the tumour sample = exhibits=20 the signature. This probability is then displayed as a heatmap with red=20 representing high probability and blue representing low probability. The = samples=20 are then clustered hierarchically on the basis of the predictions, = revealing=20 patterns of signatures across the samples.

= High=20 resolution image and legend (101 KB) =
= Figures=20 and tables index
D= ownload=20 Power Point slide (299 KB)

As a consequence of a large number of studies making use = of DNA=20 microarray analysis in cancer contexts, a diverse set of expression = signatures=20 has emerged with relevance for cancer phenotypes that can be applied in = future=20 studies. In many instances, these signatures can be directly used for = other=20 applications and studies whereas, in other instances, the data can be=20 're-utilized' for the development of signatures in specific = applications. An=20 excellent resource for identifying the available opportunities is the Oncomine cancer-profiling = database,=20 developed at the University of Michigan^{4, =}^{5, =}⁶,=20 Ann Arbor, USA.

Strategies for=20 gene-signature development

Two basic strategies have been described for the = analysis of DNA=20 microarray data¹.=20 One involves the discovery of structure in a given data set without = regard for=20 prior knowledge of the underlying biology =97 that is, no assumptions = are made=20 about what mechanisms might underlie a given gene expression profile. = This=20 approach, often referred to as 'unsupervised=20 analysis', simply uses the gene expression data to find structure in = the=20 data that can then be used to infer biologically meaningful structure. = As=20 pioneered by Brown, Botstein and colleagues⁷,=20 this approach can be an effective tool in classifying biological samples = into=20 categories that were not previously known to exist. The power of this = approach=20 can be seen in the work of Perou and colleagues, who have used = expression=20 patterns to define clinically significant subtypes of human breast = cancer^{8, =}⁹.

Conceptually, unsupervised analysis is perhaps the = simplest of the=20 tools that are applied to the analysis of gene expression data to find = patterns=20 or clusters of similarly expressed genes that can then be studied = further. There=20 are now numerous examples in which this approach has been used to = analyse gene=20 expression data, often uncovering biological complexity that was not = previously=20 appreciated, including the identification of previously unrecognized = tumour=20 subtypes^{1, =}²=20 or the refinement of tumour classification^10=
,=20^11=
,=20^12=
,=20^13=
,=20¹⁴⁼.

Although most studies that use expression data to = dissect cancer=20 phenotypes have relied on mRNA expression information, it is now clear = that the=20 expression of microRNAs (miRNAs) can be used to identify cancer-relevant = signatures^15=
,=20^16=
,=20^17=
,=20^18=
,=20¹⁹⁼.=20 Although the data content of miRNA profiles is far less than that of = mRNA=20 profiles, the regulatory nature of miRNAs suggests that this class of = gene=20 products is enriched in information content. For a recent discussion of = the role=20 of miRNAs in cancer, including the identification of cancer-relevant = signatures,=20 see Ref. 20= .

By contrast, 'supervised analysis' strategies do = consider existing=20 information and, indeed, use it to guide the analysis of the gene = expression=20 data. This approach has been particularly useful in the identification = of gene=20 expression patterns that relate to clinically relevant phenotypes such = as the=20 ability to predict the potential for recurrence of disease^21=
,=20^22=
,=20^23=
,=20^24=
,=20^25=
,=20^26=
,=20^27=
,=20^28=
,=20^29=
,=20^30=
,=20^31=
,=20^32=
,=20³³⁼.=20 The power of the supervised analysis lies in the ability to specifically = drive=20 the analysis to the phenotype of interest, taking advantage of the = relevant=20 information as a guide. Moreover, when the underlying biology that is = relevant=20 to the phenotype is uncertain, this might be the only approach to = finding those=20 gene expression patterns that relate to the phenotype. Equally = important, if the=20 clinical outcome reflects multiple components of the subtypes that are = defined=20 by unsupervised analysis, then it is only when the analysis is driven by = the=20 actual clinical outcome that this complexity can be addressed.

Much of the work using genome-scale gene expression data = has=20 focused on biological discovery, identifying genes in the profiles that = can lead=20 to a better understanding of the biological processes and, perhaps, new=20 therapeutic targets. It is logical to assume that if one can identify an = expression profile that defines a unique new subtype of breast cancer, = or one=20 that predicts which tumour is most likely to metastasize, it should be = possible=20 to identify genes in these profiles that could represent new drug = targets.=20 Nevertheless, gaining an understanding of the underlying biology from = the=20 various gene expression profiles remains a challenge. Although one can = often=20 identify function associated with some of the genes in a signature, the=20 challenge is to put this information into perspective with regard to the = entire=20 genomic profile. To appreciate this point, consider the development of a = 70-gene=20 profile that predicts breast cancer recurrence^29=
,=20³⁰⁼.=20 An examination of the genes in the profile reveals many that have = annotations=20 such as cell cycle, invasion, metastasis, angiogenesis and signal = transduction =97=20 all obviously relevant for a cancer-recurrence phenotype. Nevertheless, = they=20 represent general categorizations and reflect only a small number of the = genes=20 from the overall profile. So, by focusing on these 'recognizable' genes, = we=20 ignore many of the genes in the profile. As such, it is entirely = possible that=20 the biological and clinical deductions regarding the few recognizable = genes is=20 taken out of context. In addition, it turns out that from the same data = set=20 multiple 70-gene profiles that predict breast cancer recurrence can be=20 identified, thus expanding the number of informative genes and = emphasizing the=20 complexity of the biology that underlies this clinical phenotype³⁴⁼.

An approach that was developed to go beyond the analysis = of=20 patterns on a gene-by-gene basis, known as gene-set enrichment analysis = (GSEA),=20 uses statistical measures of enrichment of annotated gene sets within = expression=20 profiles³⁵⁼.=20 The value of GSEA and another, similar approach, 'gene module = map'³⁶⁼,=20 is to attempt to examine the true context by looking at representations = of gene=20 sets that might better reflect the underlying biology. In other words, = the power=20 lies in the ability to tap into a wide array of gene sets, including = signalling=20 pathways. Although this is a step towards putting genes into a = functional=20 context, appreciating the true meaning of the individual genes or sets = of genes=20 in a signature remains a challenge.

Using=20 cancer-relevant expression signatures

The concept of the expression signature as representing = a distinct=20 and well defined experimental state that can then be connected to an = otherwise=20 unrelated biological system opens the way for a myriad of applications = to inform=20 and understand biological complexity.

The ability of an expression signature to connect two = states,=20 where the expression signature is the intermediary, is exemplified in = the recent=20 studies of Golub and colleagues, which describe a 'Connectivity = Map'³⁷⁼.=20 In this case, expression profiles have been generated from a = well-characterized=20 cancer cell line that has been treated with many small molecule = compounds so as=20 to create a library of signatures of a drug response. This library of = signatures=20 is then used as a database that can be queried with expression = information for=20 other biological contexts, thereby linking otherwise disparate = physiological=20 events. In this context, the expression signature is represented as a = group of=20 gene identities, not by the actual properties of expression that are = defined in=20 the experimental setting. The power of this approach lies in its = independence of=20 the methodology for determining expression =97 that is, differences = between assay=20 platform or methods of measuring actual expression. Instead, a signature = is=20 simply a list of genes obtained from one experiment that is then = compared to=20 other gene lists to find matches, quantitatively assessing the degree of = similarity between the gene lists.

A conceptually similar, although mechanistically = distinct,=20 approach to the Connectivity Map makes use of quantitative aspects of = the=20 expression profile to examine other samples for similar properties. In = essence,=20 the task is to define a profile or signature experimentally and then ask = if the=20 same pattern can be recognized in another sample. Although this approach = requires that expression data are largely similar, because one is = examining the=20 extent to which the actual expression properties that are characteristic = of the=20 profile are present, this approach has the power of quantitative = assessment =97=20 generally, a probability that the signature (phenotype) is seen in the = test=20 sample. Importantly, not only does the quantitation provide a means to = estimate=20 relative contributions of signatures across samples, or the relative=20 contributions of multiple signatures in one sample, but it also provides = the=20 opportunity to identify patterns of signatures, not unlike the ability = to find=20 patterns of gene expression. Below, we provide several examples to show = how the=20 signature concept can be applied.

Identification of new therapeutic opportunities = from the=20 Connectivity Map. A search of the Connectivity Map database for = profiles=20 that coincided with a gene expression signature of glucocorticoid = sensitivity or=20 resistance in acute lymphoblastic leukaemia (ALL) cells revealed a link = with a=20 profile of the mTOR inhibitor rapamycin³⁸⁼.=20 Further experimental assays provided evidence that rapamycin could = induce=20 glucocorticoid sensitivity in lymphoid tumour cells and sensitization to = apoptosis through the modulation of the antiapoptotic gene MCL1. These findings provide a potential path = towards a new=20 therapeutic strategy involving rapamycin and glucocorticoid.

In another example, screens were carried out for = compounds that=20 inhibit a therapeutically relevant expression signature, in this case, = androgen=20 receptor (AR) = signalling. The=20 results were of limited use because the screen yielded compounds such as = celastrol and gedunin, the molecular targets of which were unknown. = However,=20 comparison of these signatures for inhibition of AR signalling against = the=20 Connectivity Map library provided connections to heat shock protein 90 = (HSP90)=20 inhibitors, suggesting HSP90 inhibition as a strategy for controlling AR = activity³⁹⁼.

Signatures that predict pathway activation in=20 cancer. One example of the use of signatures to define biology = can be=20 seen in the generation of signatures of oncogenic pathway deregulation. = Again,=20 defined experimental conditions are used to create circumstances that = represent=20 'pathway off' or 'pathway on'; expression profiles that accurately = reflect this=20 process are then developed^40=
,=20^41=
,=20^42=
,=20^43=
,=20^44=
,=20^45=
,=20⁴⁶⁼.=20 In one example, samples of quiescent cells versus RAS-, MYC-, SRC-, E2F3-, or -cate= nin-expressing=20 cells (all of which contribute to a cell proliferation response) were = used to=20 develop signatures that predict activation of these pathways⁴⁵⁼.=20 The fact that these signatures do indeed reflect pathway activation is = evident=20 because they can be used to accurately predict the molecular basis of = various=20 mouse models of cancer. Specifically, tumours arising from mice that are = transgenic for mouse mammary tumor virus (MMTV)=96MYC or MMTV=96HRAS, = and tumours=20 arising from the loss of retinoblastoma 1 (RB), were correctly = identified with=20 the appropriate pathway signature^42=
,=20^43=
,=20⁴⁵⁼.

These results demonstrate the ability of these pathway = profiles to=20 predict tumours that arise from deregulation of the corresponding = pathways, and=20 provide evidence that this approach can be used as a basis for = characterization=20 of the status of the pathways in various tumour contexts. Importantly, = the assay=20 of gene expression profiles provides a measure of the consequence of the = oncogenic process, irrespective of how the pathway might have been = altered. As=20 an example, one might revisit the classic studies of Fearon and = Vogelstein that=20 identified the accumulation of genetic alterations as a function of = colon=20 carcinoma initiation and progression from a premalignant adenomatous = stage to a=20 more aggressive carcinoma⁴⁷⁼.=20 Using pathway signatures as representations of the consequence of the=20 accumulated mutations, one might anticipate a somewhat different picture = from=20 that seen by gene mutation alone, with signatures reflecting the = consequence of=20 these mutations irrespective of how they might have been activated. For=20 instance, the RAS pathway might be active either as a result of a = mutation of=20 RAS, a mutation of a growth receptor or an alteration in one of the RAS = effector=20 activities. Moreover, the signature-based pathway analysis approach = provides an=20 opportunity to integrate these signatures to discover patterns of = pathway=20 activation, adding further value to the analysis. Indeed, predictions of = pathway=20 status in a series of lung cancer samples, followed by clustering of the = samples=20 on the basis of the oncogenic pathway signatures, has revealed distinct = patterns=20 in which subgroups of tumours are identified using pathway = patterns⁴⁵⁼.=20 Furthermore, survival analysis based on the expression patterns reveals = that the=20 ability to integrate pathway analysis by identifying patterns of pathway = deregulation does provide a way of better categorizing lung cancer = patients.

Pathway predictors have also been shown to have the = potential to=20 direct the use of therapeutics that target a component of that pathway. = In=20 particular, a measure of probability of pathway activation in a panel of = cancer=20 cell lines, on the basis of the probability that is predicted from the = pathway=20 signature, reveals a relationship between the state of pathway = deregulation and=20 sensitivity to the respective therapeutic^45=
,=20⁴⁸⁼.=20 This relationship is further emphasized by studies showing that cells = that=20 harbour a mutation in BRAF (a proto-oncogene encoding = serine/threonine-protein=20 kinase) are sensitive to inhibition of MEK (mitogen-activated protein = kinase=20 kinase 1)⁴⁹⁼.=20 Taken together, the results would suggest a potential strategy where the = prediction of pathway activation, using a collection of signatures, = could serve=20 to guide the use of targeted therapeutic agents that otherwise lack a = biomarker=20 to direct their use.

Signatures that predict drug sensitivity. = Although=20 the strategy of the Connectivity Map is to identify signatures of drug = effects,=20 an alternative approach is to develop signatures that predict drug = sensitivity.=20 The ability to predict patient response to various cancer therapies, = including=20 commonly used cytotoxic chemotherapies, is key to achieving the goal of=20 personalized medicine whereby patients are accurately matched to the = therapy or=20 therapies that best address the biology of their cancer. The logical = approach is=20 to make use of clinical studies in which patients are treated with a = given drug,=20 their response is measured and gene expression data are generated from = samples=20 that were collected before treatment in order to develop a profile that = can=20 predict response. Indeed, this strategy has already been used = successfully in=20 several studies^48=
,=20^50=
,=20⁵¹⁼.=20 The main limitation to this approach is that clinical studies are = lengthy and=20 focused on a particular regimen.

In an alternative strategy, the same goal is = accomplished by using=20 cancer cell lines grown in culture. In one such case, a panel of cancer = cell=20 lines were treated with dasatinib, a multitargeted kinase inhibitor, and = sensitivity to the drug was measured. In parallel, expression data = generated=20 from the same panel of cell lines was used to develop a signature to = predict=20 sensitivity to the drug⁵²⁼.=20 Interestingly, the cells that were defined in this manner as sensitive = largely=20 coincided with those showing evidence of SRC pathway activation, as = measured by=20 a SRC pathway signature⁴⁵⁼.=20 A second example involves the use of panels of lung cancer cell lines to = develop=20 gene expression signatures that predict sensitivity to the epidermal = growth=20 factor receptor (EGFR)=20 inhibitors gefitnib⁵³⁼=20 or erlotinib⁵⁴⁼.=20 Of these two, the erlotinib-sensitivity signature could more effectively = predict=20 sensitivity, possibly suggesting that a signature of sensitivity to=20 EGFR-directed therapy captures the consequence of many events that = contribute to=20 the phenomenon of geftinib and erlotinib sensitivity.

An analogous strategy has taken advantage of existing=20 drug-sensitivity data derived from the NCI-60=20 cancer cell line panel. This data set contains a wealth of = drug-sensitivity=20 measures coupled with baseline Affymetrix gene expression data. Cells = that were=20 identified as sensitive and resistant to various commonly used cytotoxic = chemotherapies have been used to generate expression profiles that could = predict=20 this sensitivity^55=
,=20⁵⁶⁼.=20 Importantly, multiple predictors of chemosensitivity have been shown to = also=20 predict response to the drugs in patient studies⁵⁶⁼.

Signature of chromosomal instability that predicts = cancer=20 outcome. Using the measured state of aneuploidy in a series of = tumour=20 samples, recent work has developed an expression signature that reflects = chromosomal instability⁵⁷⁼.=20 Given the variability in how aneuploidy might be measured, the signature = provides a quantitative estimate of this trait that can then be applied = to other=20 cancer contexts, including an evaluation of the extent to which this = phenotype=20 predicts clinical outcome. Indeed, the signature was found to be = elevated in=20 metastatic specimens compared with primary tumours, providing a means to = measure=20 the role of chromosomal instability in determining malignant potential, = possibly=20 in a variety of tumours.

Using signatures as phenotypes for genetic=20 discovery. In a recent study, an expression signature was used = to=20 identify genomic alterations that are tightly linked with this = 'surrogate=20 phenotype'. The signature was the so-called wound-response signature, = developed=20 experimentally on the basis of serum stimulation of fibroblasts in=20 culture⁵⁸⁼.=20 The authors analysed a series of breast tumours for evidence of the=20 wound-healing phenotype, based on the expression signature, and looked = for=20 association with DNA copy number changes⁵⁹⁼=20 (Fig= .=20 3). The method of analysis, termed stepwise linkage analysis of = microarray=20 signatures (SLAMS), is derived from the concept of genetic association = in which=20 a genetic locus that is responsible for driving a phenotype is = identified on the=20 basis of co-segregation with individuals (in this case, tumour samples) = that=20 exhibit the phenotype (in this case, an expression signature). The = authors=20 showed that the wound-response signature, which is also a poor-prognosis = expression pattern of 512 genes in breast cancer, is associated with = coordinate=20 amplifications of MYC and CSN5=20 (also known as JAB1 or COPS5). In short, the authors used a gene = expression=20 signature as a relevant breast cancer phenotype =97 or subphenotype =97 = to carry out=20 a genetic association study to identify those genomic alterations that = are=20 tightly linked to this phenotype.

= Figure=20 3 | Stepwise linkage analysis of microarray = signatures (SLAMS).

a | The use of an expression signature as a = phenotype in=20 a genetic linkage strategy to identify loci that show amplification in = relation=20 to the signature is shown. Samples are initially divided into two groups = on the=20 basis of their expression signature (b). Then, = genome copy=20 data is used to associate gene copy number changes with expression = signature.=20 Candidate genes are filtered on the basis of actual expression compared = to the=20 signature. Finally, candidate genes are further validated, on the basis = of=20 expression, in additional sample sets. Reproduced with permission from = Nature Genetics Ref. 59= =A9=20 (2006) Macmillan Publishers Ltd.

= High=20 resolution image and legend (93 KB)
= Figures=20 and tables index
D= ownload=20 Power Point slide (290 KB)

The recent identification of the MITF gene as a melanoma oncogene made use of what = is, in=20 effect, a reversal of the SLAMS strategy. In particular, SNP copy number = data=20 from the NCI-60 panel was used to identify a melanoma-specific copy = number gain=20 on chromosome 3. This information was then used to identify expression = patterns=20 coincident with the SNP profile, leading to the identification of MITF⁶⁰⁼.

Challenges of using=20 expression signatures

Although the various initial studies that are summarized = here=20 provide strong evidence for the power of the concept of using expression = signatures to define cancer phenotypes, it must be recognized that there = are=20 significant challenges inherent in this approach that reflect variation = in both=20 the biology and the technical approaches.

The molecular heterogeneity of cancer, resulting from = the=20 acquisition of multiple genetic alterations that contribute to the = development=20 of the tumour, underlies the heterogeneity of carcinogenesis. Indeed, = this=20 heterogeneity might well contribute to the observation that one can = often find=20 multiple expression profiles that predict a common phenotype such as = breast=20 cancer recurrence^34=
,=20⁶¹⁼.=20 Technical variation can also be a factor that contributes to variation = in=20 profiles that are identified in distinct studies. Platform differences = and batch=20 effects between studies have a significant impact on the ability to = consistently=20 apply expression signatures to independent validations⁶²⁼,=20 potentially limiting the usefulness of the signatures. The impact of = this=20 technical variation is reduced by the Connectivity Map strategy, because = it does=20 not depend on actual expression data but rather just the identification = of=20 genes.

Although the use of expression signatures is = conceptually simple=20 and has clearly been shown to represent important biology in many = instances, as=20 with any approach, these methodologies must be further validated in = larger=20 studies that assess the extent to which a signature is robust across = various=20 studies. Furthermore, attention to appropriate statistical design in = these=20 studies, as well as demonstration of validation in multiple independent=20 settings, will also be crucial to demonstrating the utility of using = expression=20 signatures in clinical practice to better understand cancer biology and = to=20 improve patient outcomes.

Future directions

The development of a library of expression-based = signatures that=20 can serve as surrogate phenotypes for various biological states = represents a=20 powerful tool for dissecting the complexity of biological processes. = Linking=20 biological states with specific genetic alterations and the capacity to = utilize=20 the information should guide new therapeutic strategies. The ability to=20 integrate these signatures into a coherent view of biological pathways = and=20 systems will provide an opportunity to address the complexities of = cellular=20 signalling events, particularly the interplay of various key = pathways.

Taking advantage of complexity by integration of=20 signatures. The use of expression signatures as cancer = phenotypes has=20 already been shown to have significant value in dissecting complex = biological=20 states. The challenge for the future will be to take these analyses to a = higher=20 level, integrating the various signatures, or phenotypes, into a more = complex=20 view of the biological states, including the structure and function of = cellular=20 signalling events. Moreover, the power of 'combined signatures' will = probably be=20 improved by the integration of new signatures that represent different=20 biological events such as pathway activation or chemosensitivity as = opposed to=20 what is achieved by just combining prognostic signatures. Indeed, recent = examples have highlighted the potential power of integrating diverse = signatures=20 with clinical information to yield more robust predictive models in = breast=20 cancer^32=
,=20⁶¹⁼.

Using signatures to guide the interpretation of = DNA sequence=20 variation. The Cancer = Genome=20 Atlas initiative is a recently initiated large-scale project with = the=20 ambitious goal of characterizing the DNA sequence variation that is = unique to=20 three human cancers. Perhaps the most significant challenge for the = overall=20 Cancer Genome Atlas initiative will be the ability to interpret the DNA = sequence=20 information from a collection of tumour samples =97 to identify those = sequence=20 variations that are significant with respect to a cancer phenotype. This = is=20 already illustrated by the initial observations of the cancer genome = sequencing=20 efforts: the identification of large numbers of mutations makes it = difficult to=20 interpret the role of these alterations in defining the phenotype^62=
,=20^63=
,=20^64=
,=20⁶⁵⁼.=20 We believe that one major limitation in this work is the lack of more = precise=20 phenotypes that are linked to the sequence data. Cancer is a complex = disease and=20 biological process with tremendous variation and heterogeneity, even = within a=20 single cancer type. Therefore, an ability to better describe individual=20 phenotypes that might result from a unique collection of DNA sequence=20 alterations offers the potential for better linking the sequence data = with=20 clinical phenotypes.

A potentially powerful way of addressing this issue = would be to=20 use the multitude of cancer-relevant phenotypes that are generated from=20 expression signatures together with the concept of SLAMS, as described = by Chang=20 and colleagues⁵⁹⁼.=20 In principle, this concept of linking a phenotype, as defined by a gene=20 expression signature, to a genetic alteration can be extended to the = analysis of=20 DNA sequence variation identified across cancer genomes (Fig= . 4).=20 Various cancer-relevant signatures could be used as cancer subphenotypes = to=20 drive the analysis of DNA sequence variation in a manner that is = analogous to=20 the work of Chang and colleagues. These signatures could be derived from = many=20 that already exist, reflecting diverse aspects of cancer biology such as = oncogenic signalling pathway activation, clinical outcomes and response = to=20 therapy. Alternatively, they could represent newly designed signatures = that=20 reflect other relevant aspects of cancer biology. Importantly, the = ability to=20 identify patterns of signatures, as shown by the example in Fig= . 4,=20 provides an opportunity to go even further in the use of this = information to=20 derive cancer-relevant phenotypes.

= Figure=20 4 | Use of signature patterns to identify = complex=20 cancer-relevant phenotypes to guide sequence analysis.

Sequence analysis of 1,000 selected genes in a collection of 500 = tumour=20 samples is shown. In parallel, the same tumour samples are used for gene = expression analysis and signature profiling to produce clusters of = signature=20 profiles that reflect biological phenotypes. The patterns of expression=20 signatures are then used as 'phenotypes' to guide the analysis of the = DNA=20 sequence data in a manner that is analogous to the concept of SLAMS = (stepwise=20 linkage analysis of microarray signatures; see Fig= . 3).=20 The goal is to identify DNA sequence variation in specific tumours that = coincide=20 with signature-based phenotypes.

= High=20 resolution image and legend (59 KB)
= Figures=20 and tables index
D= ownload=20 Power Point slide (256 KB)

Conclusions

The ability to use expression signatures to dissect the = complexity=20 of cancer phenotypes and ultimately to discover the mechanisms that = underlie=20 cancer have been amply demonstrated by now. The development of = expression=20 signatures that is summarized here is just the beginning =97 we envision = that many=20 more representations of important experimental and clinical biology can = be=20 utilized for the development of these powerful tools. Equally important = will be=20 an ability to understand the biological processes that are reflected in = the=20 signatures, placing the expression and function of genes in a broader = context of=20 pathways and networks. But, even before this important goal is achieved, = the=20 expression signatures have significant value as phenotypes, irrespective = of an=20 understanding of the specific genes that constitute the = profile.

Acknowledgements

We thank A. Bild for many helpful discussions and K. = Culler for=20 assistance with the preparation of the manuscript.

Competing interests statement

The authors declare no competing financial=20 interests.

To= p of page

References

Golub, T. R. et al. Molecular=20 classification of cancer: class discovery and class prediction by gene = expression monitoring. Science 286, 531=96537 (1999).
The paradigm for use of = both=20 unsupervised and supervised methods of gene expression analysis to = define new=20 classes of tumours and predict these classes in new samples.
- Article
- PubMed
- ISI
- ChemPort
Alizadeh, A. A. et al. Distinct types=20 of diffuse large B-cell lymphoma identified by gene expression=20 profiling. Nature 403,=20 503=96511 (2000).
- Article
- PubMed
- ISI
- ChemPort
Hanahan, D. & Weinberg, R. A. The=20 hallmarks of cancer. Cell 100, 57=9670 (2000).
- Article
- PubMed
- ISI
- ChemPort
Rhodes, D. R. et al. Mining for=20 regulatory programs in the cancer transcriptome. Nature=20 Genet. 37, 579=96583 (2005).
- Article
- Rhodes, D. R. & Chinnaiyan, A. M. Integrative analysis of the cancer transcriptome. = Nature Genet. 37, = S31=96S37 (2005).
  - Article
  - Rhodes, D. R. et al. Large-scale=20 meta-analysis of cancer microarray data identifies common = transcriptional=20 profiles of neoplastic transformation and progression. Proc. Natl Acad. Sci. USA 101,=20 9309=969314 (2004).
    - Article
    - PubMed
    - ChemPort
  - Eisen, M. B., Spellman, P. T., Brown, P. O. & = Botstein,=20 D. Cluster analysis and display of genome-wide = expression=20 patterns. Proc. Natl Acad. Sci. USA = 95, 14863=9614868 (1998).
    - Article
    - PubMed
    - ChemPort
  - Sorlie, T. et al. Repeated = observation=20 of breast tumor subtypes in independent gene expression data = sets.=20 Proc. Natl Acad. Sci. USA 100,=20 8418=968423 (2003).
    - Article
    - PubMed
    - ChemPort
  - Perou, C. M. et al. Molecular=20 portraits of human breast tumors. Nature 406, 747=96752 (2000).
    - Article
    - PubMed
    - ISI
    - ChemPort
  - Ramaswamy, S. et al. Multiclass cancer=20 diagnosis using tumor gene expression signatures. Proc.=20 Natl Acad. Sci. USA 98, 15149=9615154 = (2001).
    - Article
    - PubMed
    - ChemPort
  - Lapointe, J. et al. Gene = expression=20 profiling identifies clinically relevant subtypes of prostate = cancer.=20 Proc. Natl Acad. Sci. USA 101,=20 811=96816 (2004).
    - Article
    - PubMed
    - ChemPort
  - Hayes, D. N. et al. Gene = expression=20 profiling reveals reproducible human lung adenocarcinoma subtypes in = multiple=20 independent patient cohorts. J. Clin. = Oncol.=20 24, 5079=965090 (2006).
    - Article
    - PubMed
    - ISI
    - ChemPort
  - Dave, S. S. et al. Molecular diagnosis=20 of Burkitt's lymphoma. N. Engl. J. = Med. 354, 2431=962442 (2006).
    - Article
    - PubMed
    - ISI
    - ChemPort
  - Dave, S. S. et al. Prediction of=20 survival in follicular lymphoma based on molecular features of=20 tumor-infiltrating immune cells. N. Engl. J.=20 Med. 351, 2159=962169 (2004).
    - Article
    - PubMed
    - ISI
    - ChemPort
  - He, L. et al. A microRNA = polycistron=20 as a potential human oncogene. Nature = 435, 828=96833 (2005).
    - Article<= /LI>
    - PubMed
    - ISI
    - ChemPort
  - Yanaihara, N. et al. Unique microRNA=20 molecular profiles in lung cancer diagnosis and prognosis. = Cancer Cell 9, 189=96198 = (2006).
    - Article
    - PubMed
    - ChemPort
  - Lu, J. et al. MicroRNA = expression=20 profiles classify human cancers. Nature 435, 834=96838 (2005).
    - Article<= /LI>
    - PubMed
    - ISI
    - ChemPort
  - Iorio, M. V. et al. MicroRNA gene=20 expression deregulation in human breast cancer. Cancer=20 Res. 65, 7065=967070 (2005).
    - Article
    - PubMed
    - ISI
    - ChemPort
  - Volinia, S. et al. A = microRNA=20 expression signature of human solid tumors defines cancer gene = targets.=20 Proc. Natl Acad. Sci. USA 103,=20 2257=962261 (2006).
    - Article
    - PubMed
    - ChemPort
  - Calin, G. A. & Croce, C. M. MicroRNA=20 signatures in human cancers. Nature Rev. = Cancer=20 6, 857=96866 (2006).
    - Article
  - Ramaswamy, S., Ross, K. N., Lander, E. S. & = Golub, T. R.=20 A molecular signature of metastasis in primary solid = tumors. Nature Genet. 33,=20 59=9654 (2003).
    - Article
  - Berchuck, A. et al. Patterns of gene=20 expression that characterize long term survival in advanced serous = ovarian=20 cancers. Clin. Cancer Res. 11, 3686=963696 (2005).
    - Article
    - PubMed
    - ISI
    - ChemPort
  - Beer, D. G. et al. Gene-expression=20 profiles predict survival of patients with lung adenocarcinoma. = Nature Med. 8, 816=96824 = (2002).
    - Article
    - Potti, A. et al. A genomic = strategy to=20 refine prognosis in non-small cell lung carcinoma. N.=20 Engl. J. Med. 355, 570=96580 (2006).
      - Article
      - PubMed
      - ISI
      - ChemPort
    - Paik, S. et al. A = multigene assay to=20 predict recurrence of tamoxifen-treated, node-negative breast = cancer.=20 N. Engl. J. Med. 351, 2817=962826=20 (2004).
      - Article
      - PubMed
      - ISI
      - ChemPort
    - Huang, E. et al. Gene = expression=20 predictors of breast cancer outcomes. Lancet=20 361, 1590=961596 (2003).
      - Article
      - PubMed
      - ISI
      - ChemPort
    - Shipp, M. A. et al. Diffuse large=20 B-cell lymphoma outcome prediction by gene-expression profiling and = supervised=20 machine learning. Nature Med. 8, 68=9674 (2002).
      - Article
    - Pomeroy, S. L. et al. Prediction of=20 central nervous system embryonal tumour outcome based on gene=20 expression. Nature 415,=20 436=96442 (2002).
      - Article
      - PubMed
      - ISI
      - ChemPort
    - van de Vijver, M. J. et al. A=20 gene-expression signature as a predictor of survival in breast = cancer.=20 N. Engl. J. Med. 347, 1999=962009=20 (2002).
      - Article
      - PubMed
      - ISI
      - ChemPort
    - van'T Veer, L. J. et al. Gene=20 expression profiling predicts clinical outcome of breast = cancer. Nature 415, 530=96536 = (2002).
      The initial example of a = gene=20 expression profile developed to refine and improve clinical = prognosis.
      - Article
      - PubMed
      - ISI
      - ChemPort
    - Singh, D. et al. Gene = expression=20 correlates of clinical prostate cancer behavior. Cancer=20 Cell 1, 203=96209 (2002).
      - Article
      - PubMed
      - ISI
      - ChemPort
    - Pittman, J. et al. Integrated modeling=20 of clinical and gene expression information for personalized = prediction of=20 disease outcomes. Proc. Natl Acad. Sci. = USA=20 101, 8431=968436 (2004).
      - Article
      - PubMed
      - ChemPort
    - West, M. et al. Predicting = the=20 clinical status of human breast cancer by using gene expression=20 profiles. Proc. Natl Acad. Sci. USA = 98, 11462=9611467 (2001).
      - Article
      - PubMed
      - ChemPort
    - Ein-Dor, L., Kela, I., Getz, G., Givol, D. & = Domany, E.=20 Outcome signature genes in breast cancer: is there a = unique=20 set? Bioinformatics 21,=20 171=96178 (2005).
      - PubMed
      - ISI
      - ChemPort
    - Subramanian, A. et al. Gene set=20 enrichment analysis: a knowledge-based approach for interpreting = genome-wide=20 expression profiles. Proc. Natl Acad. Sci. = USA=20 102, 15545=9615550 (2005).
      A description of GSEA = methodology as=20 a tool to identify biological context in gene expression = profiles.
      - Article
      - PubMed
      - ChemPort
    - Segal, E., Friedman, N., Koller, D. & Regev, A. = A module map showing conditional activity of expression = modules in=20 cancer. Nature Genet. 36,=20 1090=961098 (2004).
    - Lamb, J. et al. The = Connectivity Map:=20 using gene expression signatures to connect small molecules, genes, = and=20 disease. Science 313,=20 1929=961935 (2006).
      A = description of a=20 novel approach to connect two biological states, using gene expression = as the=20 intermediary.
      - Article
      - PubMed
      - ISI
      - ChemPort
    - Wei, G. et al. Gene = expression based=20 chemical genomics identifies rapamycin as a modulator of MCL1 and=20 glucocorticoid resistance. Cancer Cell = 10, 331=96342 (2006).
      - Article
      - PubMed
      - ISI
      - ChemPort
    - Hieronymus, H. et al. Gene = expression=20 signature-based chemical genomic prediction identifies a novel class = of HSP90=20 pathway modulators. Cancer Cell 10, 321=96330 (2006).
      - Article
      - PubMed
      - ISI
      - ChemPort
    - Desai, K. V. et al. Initiating=20 oncogenic event determines gene-expression patterns of human breast = cancer=20 models. Proc. Natl Acad. Sci. USA = 99, 6967=966972 (2002).
      - Article
      - PubMed
      - ChemPort
    - Ferrando, A. A. et al. Gene expression=20 signatures define novel oncogenic pathways in T cell acute = lymphoblastic=20 leukemia. Cancer Cell 1,=20 75=9687 (2002).
      - Article
      - PubMed
      - ISI
      - ChemPort
    - Huang, E. et al. Gene = expression=20 phenotypic models that predict the activity of oncogenic = pathways.=20 Nature Genet. 34, = 226=96230 (2003).
      - Article
      - Black, E. P. et al. Distinct gene=20 expression phenotypes of cells lacking RB and RB family = members. Cancer Res. 63, = 3716=963723 (2003).
        
        PubMed
        
        ISI
        
        ChemPort
      - Sweet-Cordero, A. et al. An oncogenic=20 KRAS2 expression signature identified by cross-species gene expression = analysis. Nature Genet. 37, 48=9654 (2005).
        
        Article
        <= /UL>
        
        Bild, A. et al. Oncogenic = pathway=20 signatures in human cancers as a guide to targeted therapies. = Nature 439, 353=96357 = (2006).
        The development of gene = expression=20 signatures that reflect the activation or deregulation of various = oncogenic=20 signalling pathways, together with the utilization of these signatures = to=20 predict sensitivity to drugs that target the pathways.
        
        Article<= /LI>
        PubMed
        
        ISI
        
        ChemPort
        
        Lamb, J. et al. A = mechanism of cyclin=20 D1 action encoded in the patterns of gene expression in human = cancer.=20 Cell 114, 323=96334 = (2003).
        
        Article
        =
        PubMed
        
        ISI
        
        ChemPort
        
        Fearon, E. R. & Vogelstein, B. A genetic=20 model for colorectal tumorigenesis. Cell 17, 671=96674 (1990).
        The first comprehensive = description=20 of genetic events involved in colon carcinogenesis.
        
        Dressman, H. K. et al. An = integrated=20 genomic-based approach to individualized treatment of patients with = advanced=20 stage ovarian cancer. J. Clin. Oncol. = 25, 517=96525 (2007).
        
        Article
        
        PubMed
        
        ISI
        
        ChemPort
        
        Solit, D. B. et al. BRAF = mutation=20 predicts sensitivity to MEK inhibition. Nature=20 439, 274=96275 (2006).
        
        Article
        =
        PubMed
        
        ChemPort
        
        Chang, J. C. et al. Gene = expression=20 profiling for the prediction of therapeutic response to docetaxel in = patients=20 with breast cancer. Lancet 362, 362=96369 (2003).
        
        Article
        =
        PubMed
        
        ISI
        
        ChemPort
        
        Ayers, M. et al. Gene = expression=20 profiles predict complete pathologic response to neoadjuvant = paclitaxel and=20 fluorouracil, doxorubicin, and cyclophosphamide chemotherapy in breast = cancer. J. Clin. Oncol. 22, 2284=962293 (2004).
        
        Article
        
        PubMed
        
        ISI
        
        ChemPort
        
        Huang, F. et al. Identification of=20 candidate molecular markers predicting sensitivity in solid tumors to=20 dasatinib: rationale for patient selection. Cancer=20 Res. 67, 2226=962238 (2007).
        
        Article
        =
        PubMed
        
        ISI
        
        ChemPort
        
        Coldren, C. D. et al. Baseline gene=20 expression predicts sensitivity to gefitinib in non-small cell lung = cancer=20 cell lines. Mol. Cancer Res. 4, 521=96528 (2006).
        
        Article
        =
        PubMed
        
        ISI
        
        ChemPort
        
        Balko, J. M. et al. Gene = expression=20 patterns that predict sensitivity to epidermal growth factor receptor = tyrosine=20 kinase inhibitors in lung cancer cell lines and human lung = tumors.=20 BMC Genomics 7, 289 = (2006).
        
        Article
        
        PubMed
        
        ChemPort
        
        Staunton, J. E. et al. Chemosensitivity prediction by transcriptional = profiling.=20 Proc. Natl Acad. Sci. USA 98,=20 10787=9619792 (2001).
        
        Article
        
        PubMed
        
        ChemPort
        
        Potti, A. et al. A genomic = strategy to=20 guide the use of chemotherapeutic drugs in solid tumors. Nature Med. 12, = 1294=961300 (2006).
        
        Article
        <= /UL>
        
        Carter, S. L., Eklund, A. C., Kohane, I. S., = Harris, L. N.=20 & Szallasi, Z. A signature of chromosomal = instability=20 inferred from gene expression profiles predicts clinical outcome in = multiple=20 human cancers. Nature Genet. 38, 973=96974 (2006).
        
        Article
        
        Chang, H. Y. et al. Gene = expression=20 signature of fibroblast serum response predicts human cancer = progression:=20 similarities between tumors and wounds. PLoS=20 Biol. 2, 206=96214 (2004).
        
        Article
        
        ISI
        
        ChemPort
        
        Adler, A. S. et al. Genetic regulators=20 of large-scale transcriptional signatures in cancer. Nature Genet. 38, = 421=96430 (2006).
        A demonstration of the use = of an=20 expression signature as a cancer phenotype in a genetic association = study to=20 identify chromosomal alterations that associate with the = phenotype.
        
        Article
        <= /UL>
        
        Garraway, L. A. et al. Integrative=20 genomic analyses identify MITF as a lineage survival oncogene = amplified in=20 malignant melanoma. Nature 436, 117=96122 (2005).
        
        Article<= /LI>
        PubMed
        
        ISI
        
        ChemPort
        
        Fan, C. et al. Different = gene=20 expression based predictors for breast cancer patients are = concordant.=20 N. Engl. J. Med. 355, 560=96569=20 (2006).
        
        Article
        
        PubMed
        
        ISI
        
        ChemPort
        
        The MicroArray Quality Control Consortium. The MicroArray Quality Control (MAQC) project shows inter- = and=20 intraplatform reproducibility of gene expression measurements. = Nature Biotechnol. 24, = 1151=961161=20 (2006).
        
        Article
        =
        
        Chang, H. Y. et al. Robustness,=20 scalability, and integration of a wound-response gene expression = signature in=20 predicting breast cancer survival. Proc. Natl = Acad.=20 Sci. USA 102, 3738=963743 (2005).
        
        Article
        
        PubMed
        
        ChemPort
        
        Davies, H. et al. Somatic = mutations of=20 the protein kinase gene family in human lung cancer. Cancer Res. 65, = 7591=967595 (2005).
        
        PubMed
        
        ISI
        
        ChemPort
        
        Greenman, C. et al. Patterns of=20 somatic mutation in human cancer genomes. Nature=20 446, 153=96158 (2007).
        
        Article<= /LI>
        PubMed
        
        ISI
        
        ChemPort
        
        Parsons, D. W. et al. Colorectal=20 cancer: mutations in a signalling pathway. Nature 436, 792 (2005).
        
        Article
        =
        PubMed
        
        ISI
        
        ChemPort
        
        Sjoblom, T. et al. The = consensus=20 coding sequences of human breast and colorectal cancers. Science 314, 268=96274 = (2006).
        
        Article
        
        PubMed
        
        ISI
        
        ChemPort

To= p of page

Author affiliations

Duke Institute for Genome Sciences and Policy, Duke = University=20 Medical Center, Durham, North Carolina 27710, USA.

Correspondence to: Joseph R. Nevins¹=20 Email: nevin001@mc.duke.edu

To= p of page

Main navigation

NPG services

Help=20
Authors and = Referees=20
Librarian=20 gateway=20
Advertising=20 information=20
work@npg=20
Reprints and = permissions=20

NPG=20 resources

NPG = Journals

by Subject Area

Chemistry
- Chemistry=20
- Drug = discovery=20
- Biotechnology=20
- Materials=20
- Methods = &=20 Protocols
Clinical Practice & Research
- Cancer=20
- Cardiovascular=20 medicine=20
- Dentistry=20
- Endocrinology= =20
- Gastroenterology=20 & Hepatology=20
- Methods = &=20 Protocols=20
- Pathology = &=20 Pathobiology=20
- Urology =
Earth & Environment
- Earth = sciences=20
- Evolution = &=20 Ecology
Life sciences
- Biotechnology=20
- Cancer=20
- Development=20
- Drug = discovery=20
- Evolution = &=20 Ecology=20
- Genetics=20
- Immunology=20
- Medical=20 research=20
- Methods = &=20 Protocols=20
- Microbiology=20
- Molecular cell=20 biology=20
- Neuroscience=20
- Pharmacology=20
- Systems = biology=20
Physical sciences
- Physics =
- Materials=20

by A - Z=20 Index

Extra navigation

Subscribe to Nature Reviews = Genetics

This issue

Article tools

Article navigation

Article view

At= a=20 glance
Gl= ossary
= Figures=20 and tablesMore articles like = this

Search Pub Med for = the=20 following terms:

Joseph=20 R. Nevins
Anil=20 Potti

nature jobs

Molecular Lab = Tech=20 position (PCR, Sub-Cloning, sequencing) experience needed at the = NIH!
- KSR=20
- Rockville, MD, USA
Clinical Study=20 Coordinator for Med Device Compay in MN - Direct Hire
- KSR=20
- Sterling Heights, MI, USA

More science = jobs

To= p

Nature Reviews Genetics

ISSN: 1471-0056

EISSN: 1471-0064

=A9 2008 = Nature=20 Publishing Group =96 partner of AGORA, HINARI, = CrossRef and=20 COUNTER

").append(res.responseText.replace(//g,"")).find(selector):res.responseText);setTimeout(function()= {self.each(callback,[res.responseText,status,res]);},13);}});return = this;},serialize:function(){return = jQuery.param(this.serializeArray());},serializeArray:function(){return = this.map(function(){return = jQuery.nodeName(this,"form")?jQuery.makeArray(this.elements):this;}).filt= er(function(){return = this.name&&!this.disabled&&(this.checked||/select|textarea/i.test(this.no= deName)||/text|hidden|password/i.test(this.type));}).map(function(i,elem)= {var val=3DjQuery(this).val();return = val=3D=3Dnull?null:val.constructor=3D=3DArray?jQuery.map(val,function(val= ,i){return{name:elem.name,value:val};}):{name:elem.name,value:val};}).get= ();}});jQuery.each("ajaxStart,ajaxStop,ajaxComplete,ajaxError,ajaxSuccess= ,ajaxSend".split(","),function(i,o){jQuery.fn[o]=3Dfunction(f){return = this.bind(o,f);};});var jsc=3D(new = Date).getTime();jQuery.extend({get:function(url,data,callback,type){if(jQ= uery.isFunction(data)){callback=3Ddata;data=3Dnull;}return = jQuery.ajax({type:"GET",url:url,data:data,success:callback,dataType:type}= );},getScript:function(url,callback){return = jQuery.get(url,null,callback,"script");},getJSON:function(url,data,callba= ck){return = jQuery.get(url,data,callback,"json");},post:function(url,data,callback,ty= pe){if(jQuery.isFunction(data)){callback=3Ddata;data=3D{};}return = jQuery.ajax({type:"POST",url:url,data:data,success:callback,dataType:type= });},ajaxSetup:function(settings){jQuery.extend(jQuery.ajaxSettings,setti= ngs);},ajaxSettings:{global:true,type:"GET",timeout:0,contentType:"applic= ation/x-www-form-urlencoded",processData:true,async:true,data:null},lastM= odified:{},ajax:function(s){var = jsonp,jsre=3D/=3D(\?|%3F)/g,status,data;s=3DjQuery.extend(true,s,jQuery.e= xtend(true,{},jQuery.ajaxSettings,s));if(s.data&&s.processData&&typeof = s.data!=3D"string")s.data=3DjQuery.param(s.data);if(s.dataType=3D=3D"json= p"){if(s.type.toLowerCase()=3D=3D"get"){if(!s.url.match(jsre))s.url+=3D(s= .url.match(/\?/)?"&":"?")+(s.jsonp||"callback")+"=3D?";}else = if(!s.data||!s.data.match(jsre))s.data=3D(s.data?s.data+"&":"")+(s.jsonp|= |"callback")+"=3D?";s.dataType=3D"json";}if(s.dataType=3D=3D"json"&&(s.da= ta&&s.data.match(jsre)||s.url.match(jsre))){jsonp=3D"jsonp"+jsc++;if(s.da= ta)s.data=3Ds.data.replace(jsre,"=3D"+jsonp);s.url=3Ds.url.replace(jsre,"= =3D"+jsonp);s.dataType=3D"script";window[jsonp]=3Dfunction(tmp){data=3Dtm= p;success();complete();window[jsonp]=3Dundefined;try{delete = window[jsonp];}catch(e){}};}if(s.dataType=3D=3D"script"&&s.cache=3D=3Dnul= l)s.cache=3Dfalse;if(s.cache=3D=3D=3Dfalse&&s.type.toLowerCase()=3D=3D"ge= t")s.url+=3D(s.url.match(/\?/)?"&":"?")+"_=3D"+(new = Date()).getTime();if(s.data&&s.type.toLowerCase()=3D=3D"get"){s.url+=3D(s= .url.match(/\?/)?"&":"?")+s.data;s.data=3Dnull;}if(s.global&&!jQuery.acti= ve++)jQuery.event.trigger("ajaxStart");if(!s.url.indexOf("http")&&s.dataT= ype=3D=3D"script"){var = head=3Ddocument.getElementsByTagName("head")[0];var = script=3Ddocument.createElement("script");script.src=3Ds.url;if(!jsonp&&(= s.success||s.complete)){var = done=3Dfalse;script.onload=3Dscript.onreadystatechange=3Dfunction(){if(!d= one&&(!this.readyState||this.readyState=3D=3D"loaded"||this.readyState=3D= =3D"complete")){done=3Dtrue;success();complete();head.removeChild(script)= ;}};}head.appendChild(script);return;}var requestDone=3Dfalse;var = xml=3Dwindow.ActiveXObject?new ActiveXObject("Microsoft.XMLHTTP"):new = XMLHttpRequest();xml.open(s.type,s.url,s.async);if(s.data)xml.setRequestH= eader("Content-Type",s.contentType);if(s.ifModified)xml.setRequestHeader(= "If-Modified-Since",jQuery.lastModified[s.url]||"Thu, 01 Jan 1970 = 00:00:00 = GMT");xml.setRequestHeader("X-Requested-With","XMLHttpRequest");if(s.befo= reSend)s.beforeSend(xml);if(s.global)jQuery.event.trigger("ajaxSend",[xml= ,s]);var = onreadystatechange=3Dfunction(isTimeout){if(!requestDone&&xml&&(xml.ready= State=3D=3D4||isTimeout=3D=3D"timeout")){requestDone=3Dtrue;if(ival){clea= rInterval(ival);ival=3Dnull;}status=3DisTimeout=3D=3D"timeout"&&"timeout"= ||!jQuery.httpSuccess(xml)&&"error"||s.ifModified&&jQuery.httpNotModified= (xml,s.url)&&"notmodified"||"success";if(status=3D=3D"success"){try{data=3D= jQuery.httpData(xml,s.dataType);}catch(e){status=3D"parsererror";}}if(sta= tus=3D=3D"success"){var = modRes;try{modRes=3Dxml.getResponseHeader("Last-Modified");}catch(e){}if(= s.ifModified&&modRes)jQuery.lastModified[s.url]=3DmodRes;if(!jsonp)succes= s();}else=0A= jQuery.handleError(s,xml,status);complete();if(s.async)xml=3Dnull;}};if(s= .async){var = ival=3DsetInterval(onreadystatechange,13);if(s.timeout>0)setTimeout(funct= ion(){if(xml){xml.abort();if(!requestDone)onreadystatechange("timeout");}= },s.timeout);}try{xml.send(s.data);}catch(e){jQuery.handleError(s,xml,nul= l,e);}if(!s.async)onreadystatechange();return xml;function = success(){if(s.success)s.success(data,status);if(s.global)jQuery.event.tr= igger("ajaxSuccess",[xml,s]);}function = complete(){if(s.complete)s.complete(xml,status);if(s.global)jQuery.event.= trigger("ajaxComplete",[xml,s]);if(s.global&&!--jQuery.active)jQuery.even= t.trigger("ajaxStop");}},handleError:function(s,xml,status,e){if(s.error)= s.error(xml,status,e);if(s.global)jQuery.event.trigger("ajaxError",[xml,s= ,e]);},active:0,httpSuccess:function(r){try{return!r.status&&location.pro= tocol=3D=3D"file:"||(r.status>=3D200&&r.status<300)||r.status=3D=3D304||j= Query.browser.safari&&r.status=3D=3Dundefined;}catch(e){}return = false;},httpNotModified:function(xml,url){try{var = xmlRes=3Dxml.getResponseHeader("Last-Modified");return = xml.status=3D=3D304||xmlRes=3D=3DjQuery.lastModified[url]||jQuery.browser= .safari&&xml.status=3D=3Dundefined;}catch(e){}return = false;},httpData:function(r,type){var = ct=3Dr.getResponseHeader("content-type");var = xml=3Dtype=3D=3D"xml"||!type&&ct&&ct.indexOf("xml")>=3D0;var = data=3Dxml?r.responseXML:r.responseText;if(xml&&data.documentElement.tagN= ame=3D=3D"parsererror")throw"parsererror";if(type=3D=3D"script")jQuery.gl= obalEval(data);if(type=3D=3D"json")data=3Deval("("+data+")");return = data;},param:function(a){var = s=3D[];if(a.constructor=3D=3DArray||a.jquery)jQuery.each(a,function(){s.p= ush(encodeURIComponent(this.name)+"=3D"+encodeURIComponent(this.value));}= );else=0A= for(var j in = a)if(a[j]&&a[j].constructor=3D=3DArray)jQuery.each(a[j],function(){s.push= (encodeURIComponent(j)+"=3D"+encodeURIComponent(this));});else=0A= s.push(encodeURIComponent(j)+"=3D"+encodeURIComponent(a[j]));return = s.join("&").replace(/%20/g,"+");}});jQuery.fn.extend({show:function(speed= ,callback){return = speed?this.animate({height:"show",width:"show",opacity:"show"},speed,call= back):this.filter(":hidden").each(function(){this.style.display=3Dthis.ol= dblock?this.oldblock:"";if(jQuery.css(this,"display")=3D=3D"none")this.st= yle.display=3D"block";}).end();},hide:function(speed,callback){return = speed?this.animate({height:"hide",width:"hide",opacity:"hide"},speed,call= back):this.filter(":visible").each(function(){this.oldblock=3Dthis.oldblo= ck||jQuery.css(this,"display");if(this.oldblock=3D=3D"none")this.oldblock= =3D"block";this.style.display=3D"none";}).end();},_toggle:jQuery.fn.toggl= e,toggle:function(fn,fn2){return = jQuery.isFunction(fn)&&jQuery.isFunction(fn2)?this._toggle(fn,fn2):fn?thi= s.animate({height:"toggle",width:"toggle",opacity:"toggle"},fn,fn2):this.= each(function(){jQuery(this)[jQuery(this).is(":hidden")?"show":"hide"]();= });},slideDown:function(speed,callback){return = this.animate({height:"show"},speed,callback);},slideUp:function(speed,cal= lback){return = this.animate({height:"hide"},speed,callback);},slideToggle:function(speed= ,callback){return = this.animate({height:"toggle"},speed,callback);},fadeIn:function(speed,ca= llback){return = this.animate({opacity:"show"},speed,callback);},fadeOut:function(speed,ca= llback){return = this.animate({opacity:"hide"},speed,callback);},fadeTo:function(speed,to,= callback){return = this.animate({opacity:to},speed,callback);},animate:function(prop,speed,e= asing,callback){var opt=3DjQuery.speed(speed,easing,callback);return = this[opt.queue=3D=3D=3Dfalse?"each":"queue"](function(){opt=3DjQuery.exte= nd({},opt);var hidden=3DjQuery(this).is(":hidden"),self=3Dthis;for(var p = in = prop){if(prop[p]=3D=3D"hide"&&hidden||prop[p]=3D=3D"show"&&!hidden)return= = jQuery.isFunction(opt.complete)&&opt.complete.apply(this);if(p=3D=3D"heig= ht"||p=3D=3D"width"){opt.display=3DjQuery.css(this,"display");opt.overflo= w=3Dthis.style.overflow;}}if(opt.overflow!=3Dnull)this.style.overflow=3D"= hidden";opt.curAnim=3DjQuery.extend({},prop);jQuery.each(prop,function(na= me,val){var e=3Dnew = jQuery.fx(self,opt,name);if(/toggle|show|hide/.test(val))e[val=3D=3D"togg= le"?hidden?"show":"hide":val](prop);else{var = parts=3Dval.toString().match(/^([+-]=3D)?([\d+-.]+)(.*)$/),start=3De.cur(= true)||0;if(parts){var = end=3DparseFloat(parts[2]),unit=3Dparts[3]||"px";if(unit!=3D"px"){self.st= yle[name]=3D(end||1)+unit;start=3D((end||1)/e.cur(true))*start;self.style= [name]=3Dstart+unit;}if(parts[1])end=3D((parts[1]=3D=3D"-=3D"?-1:1)*end)+= start;e.custom(start,end,unit);}else=0A= e.custom(start,val,"");}});return = true;});},queue:function(type,fn){if(jQuery.isFunction(type)){fn=3Dtype;t= ype=3D"fx";}if(!type||(typeof type=3D=3D"string"&&!fn))return = queue(this[0],type);return = this.each(function(){if(fn.constructor=3D=3DArray)queue(this,type,fn);els= e{queue(this,type).push(fn);if(queue(this,type).length=3D=3D1)fn.apply(th= is);}});},stop:function(){var timers=3DjQuery.timers;return = this.each(function(){for(var = i=3D0;i-10000?r:parseFloat(jQuery.css(this.elem,this.prop))||0;},custom:fun= ction(from,to,unit){this.startTime=3D(new = Date()).getTime();this.start=3Dfrom;this.end=3Dto;this.unit=3Dunit||this.= unit||"px";this.now=3Dthis.start;this.pos=3Dthis.state=3D0;this.update();= var self=3Dthis;function t(){return = self.step();}t.elem=3Dthis.elem;jQuery.timers.push(t);if(jQuery.timers.le= ngth=3D=3D1){var timer=3DsetInterval(function(){var = timers=3DjQuery.timers;for(var = i=3D0;ithis.options.duration+this.startTime){this.now=3Dt= his.end;this.pos=3Dthis.state=3D1;this.update();this.options.curAnim[this= .prop]=3Dtrue;var done=3Dtrue;for(var i in = this.options.curAnim)if(this.options.curAnim[i]!=3D=3Dtrue)done=3Dfalse;i= f(done){if(this.options.display!=3Dnull){this.elem.style.overflow=3Dthis.= options.overflow;this.elem.style.display=3Dthis.options.display;if(jQuery= .css(this.elem,"display")=3D=3D"none")this.elem.style.display=3D"block";}= if(this.options.hide)this.elem.style.display=3D"none";if(this.options.hid= e||this.options.show)for(var p in = this.options.curAnim)jQuery.attr(this.elem.style,p,this.options.orig[p]);= }if(done&&jQuery.isFunction(this.options.complete))this.options.complete.= apply(this.elem);return false;}else{var = n=3Dt-this.startTime;this.state=3Dn/this.options.duration;this.pos=3DjQue= ry.easing[this.options.easing||(jQuery.easing.swing?"swing":"linear")](th= is.state,n,0,1,this.options.duration);this.now=3Dthis.start+((this.end-th= is.start)*this.pos);this.update();}return = true;}};jQuery.fx.step=3D{scrollLeft:function(fx){fx.elem.scrollLeft=3Dfx= .now;},scrollTop:function(fx){fx.elem.scrollTop=3Dfx.now;},opacity:functi= on(fx){jQuery.attr(fx.elem.style,"opacity",fx.now);},_default:function(fx= ){fx.elem.style[fx.prop]=3Dfx.now+fx.unit;}};jQuery.fn.offset=3Dfunction(= ){var = left=3D0,top=3D0,elem=3Dthis[0],results;if(elem)with(jQuery.browser){var = absolute=3DjQuery.css(elem,"position")=3D=3D"absolute",parent=3Delem.pare= ntNode,offsetParent=3Delem.offsetParent,doc=3Delem.ownerDocument,safari2=3D= safari&&parseInt(version)<522;if(elem.getBoundingClientRect){box=3Delem.g= etBoundingClientRect();add(box.left+Math.max(doc.documentElement.scrollLe= ft,doc.body.scrollLeft),box.top+Math.max(doc.documentElement.scrollTop,do= c.body.scrollTop));if(msie){var = border=3DjQuery("html").css("borderWidth");border=3D(border=3D=3D"medium"= ||jQuery.boxModel&&parseInt(version)>=3D7)&&2||border;add(-border,-border= );}}else{add(elem.offsetLeft,elem.offsetTop);while(offsetParent){add(offs= etParent.offsetLeft,offsetParent.offsetTop);if(mozilla&&/^t[d|h]$/i.test(= parent.tagName)||!safari2)border(offsetParent);if(safari2&&!absolute&&jQu= ery.css(offsetParent,"position")=3D=3D"absolute")absolute=3Dtrue;offsetPa= rent=3DoffsetParent.offsetParent;}while(parent.tagName&&!/^body|html$/i.t= est(parent.tagName)){if(!/^inline|table-row.*$/i.test(jQuery.css(parent,"= display")))add(-parent.scrollLeft,-parent.scrollTop);if(mozilla&&jQuery.c= ss(parent,"overflow")!=3D"visible")border(parent);parent=3Dparent.parentN= ode;}if(safari2&&absolute)add(-doc.body.offsetLeft,-doc.body.offsetTop);}= results=3D{top:top,left:left};}return results;function = border(elem){add(jQuery.css(elem,"borderLeftWidth"),jQuery.css(elem,"bord= erTopWidth"));}function = add(l,t){left+=3DparseInt(l)||0;top+=3DparseInt(t)||0;}};})(); ------=_NextPart_000_0147_01C8B204.881AA500 Content-Type: application/octet-stream Content-Transfer-Encoding: quoted-printable Content-Location: http://www.nature.com/common/scripts/global.js (function($) { $(document).ready(function () { // enables dropdown ads to display whilst maintaining .user-services = visibility aat small resolutions // yes, it involves browser detection, and yes, that is truly = horrible. if (jQuery.browser.mozilla) { =09 var dropdownAd =3D false; =09 // for each descendant of .leaderboard that isn't a

Review

Mining gene expression profiles: expression signatures as = cancer=20 phenotypes

Abstract

The = concept of gene=20 expression signatures

Developing a useful=20 signature

= Figure=20 1 | Generation of an expression = signature.

= Figure=20 2 | Identification of patterns of=20 signatures.

Strategies for=20 gene-signature development

Using=20 cancer-relevant expression signatures

= Figure=20 3 | Stepwise linkage analysis of microarray = signatures (SLAMS).

Challenges of using=20 expression signatures

Future directions

= Figure=20 4 | Use of signature patterns to identify = complex=20 cancer-relevant phenotypes to guide sequence analysis.

Conclusions

Links

FURTHER INFORMATION

DATABASES

Entrez-G= ene

UniProtKB

Acknowledgements

Competing interests statement

References

Author affiliations

MORE ARTICLES LIKE THIS

NEWS AND VIEWS

RESEARCH

Main navigation

Journal content

Journal information

NPG services

NPG=20 resources

NPG = Journals

by Subject Area

Chemistry

Clinical Practice & Research

Earth & Environment

Life sciences

Physical sciences

by A - Z=20 Index

Extra navigation

Subscribe

This issue

Article tools

Article navigation

Article view

Search Pub Med for = the=20 following terms:

nature jobs

Molecular Lab = Tech=20 position (PCR, Sub-Cloning, sequencing) experience needed at the = NIH!

Clinical Study=20 Coordinator for Med Device Compay in MN - Direct Hire

nature products