Predictive and prognostic molecular markers for cancer medicine

Abstract

Over the last 10 years there has been an explosion of information about the molecular biology of cancer. A challenge in oncology is to translate this information into advances in patient care. While there are well-formed routes for translating new molecular information into drug therapy, the routes for translating new information into sensitive and specific diagnostic, prognostic and predictive tests are still being developed. Similarly, the science of using tumor molecular profiles to select clinical trial participants or to optimize therapy for individual patients is still in its infancy. This review will summarize the current technologies for predicting treatment response and prognosis in cancer medicine, and outline what the future may hold. It will also highlight the potential importance of methods that can integrate molecular, histopathological and clinical information into a synergistic understanding of tumor progression. While these possibilities are without doubt exciting, significant challenges remain if we are to implement them with a strong evidence base in a widely available and cost-effective manner.

Introduction

Technological advances have greatly increased our understanding of the molecular basis of tumor progression and treatment response. Over the last 10 years these advances have led to the identification of numerous tumor biomarkers. These biomarkers can be divided into two types. Prognostic markers (see examples in Table 1) aim to objectively evaluate the patient’s overall outcome, such as the probability of cancer recurrence after standard treatment. The presence or absence of a prognostic marker can be useful for the selection of patients for treatment but does not directly predict the response to a treatment. Predictive markers (see examples in Table 2) aim to objectively evaluate the likelihood of benefit from a specific clinical intervention, or the differential outcomes of two or more interventions, including toxicity. This is a rapidly accelerating research field that is beginning to have a significant clinical impact. Since the year 2000, there have been over 26,000 publications indexed in PubMed with the joint medical subject headings of ‘neoplasm’ and ‘predictive marker’, and almost 14,000 publications with ‘neoplasm’ and ‘prognostic marker’. Increasingly, clinicians need to interpret molecular biomarkers and understanding the technologies that underlie them in order to make treatment decisions.

Table 1.

Prognostic biomarkers for cancer medicine.

Prognostic biomarker	Type of cancer	Clinical significance	Detection	Clinical use	References
Beta-tubulin	NSCLC	High expression of beta-tubulin confers worse prognosis	IHC	No	[Seve et al. 2007;  Winton et al. 2005]
BRCA1	Breast	High expression of BRCA1 confers worse prognosis in untreated patients	IHC	Yes	[James et al. 2007]
	NSCLC	High expression of BRCA1 confers worse prognosis in untreated patients	qRT-PCR	No	[Rosell et al. 2007]
CA19-9	Pancreatic	Higher preoperative CA19-9 levels are associated with lower resectability,  more advanced stage and inferior survival	IHC	No	[Ferrone et al. 2006]
CAIX	RCC	High expression of CAIX is associated with a better prognosis	IHC	No	[Bui et al. 2003]
CD44	Bladder	Expression of CD44 is associated with poor prognosis	qRT-PCR	No	[Miyake et al. 2002]
CEA	CRC	Elevated preoperative CEA levels in resectable colorectal cancer is  associated with poor prognosis	IHC	Yes	[Wolmark et al. 1984;  Wanebo et al. 1978]
c-KIT	GIST	GIST patients have a better prognosis if they harbor a mutation in exon 11 of  the c-KIT gene	Pathway  detection  via FDG-PET	Yes	[Singer et al. 2002]
ColoPrint	CRC	Prognosis for colorectal cancer patients	Microarray	Yes	[Glas et al. 2009]
CTC (e.g. CellSearch)	Melanoma	Increased number of circulating melanoma cells is associated with poor  prognosis	Circulating  tumor cells	No	[Nezos et al. 2009]
	CRC	Colorectal patients with ≥3 CTC/7.5 ml of peripheral blood were  associated with shorter PFS and OS, i.e. poor prognosis	Circulating  tumor cells	Yes	[Cohen et al. 2009]
	Breast	Breast cancer patients with ≥5 CTC/7.5 ml of peripheral blood are associated  with shorter PFS and OS, i.e. poor prognosis	Circulating  tumor cells	Yes	[Cristofanilli  et al. 2004]
	Prostate	≥5 CTC/7.5 ml of peripheral blood is associated with poor prognosis	Circulating  tumor cells	Yes	[Cristofanilli  et al. 2004]
Cyclin D1	Bladder	Expression of Cyclin D1 is associated with low grade, low stage  and recurrence	IHC	No	[Liukkonen et al. 2000]
Cyclin E	Bladder	Expression of Cyclin E is associated with low stage and survival	IHC	No	[Richter et al. 2000]
E-Cadherin	Bladder	E-Cadherin is associated with poor prognosis	IHC	No	[Bringuier et al. 1993]
EGFR	Bladder	Overexpression of EGFR is associated with high grade and high stage	IHC	No	[Nguyen et al. 1994]
	NSCLC	High gene copy number of EGFR in NSCLC patients is associated with  poor prognosis	FISH / SA	No	[Coate et al. 2009]
	NSCLC	EGFR mutation in NSCLC patients is associated with better prognosis  in untreated patients
	Rectal	Overexpression of EGFR in rectal cancers is also associated with  poor prognosis	IHC	No	[Kuremsky et al. 2009]
ER	Breast	Patients with ER-positive breast tumors have better survival than  patients with hormonal negative tumors	IHC	Yes	[Early Breast Cancer  Trialists’ Collaborative  Group, 1998]
eXageneBC	Breast	Provides prognosis in node-positive or node-negative breast cancer patients	FISH	Yes	[Davis et al. 2007]
Her2/neu	Breast	Patients with Her2/neu-positive breast tumors are more aggressive and  have a worse prognosis compared to Her2/neu-negative tumors	FISH	Yes	[Mass et al. 2005]
	Bladder	Overexpression of Her2/neu is associated with high grade, high stage,  poor survival and metastasis in bladder cancer	IHC	No	[Lipponen et al. 1991]
	GIST	Overexpression of Her2/neu in advanced gastric cancer patients is  associated with poor prognosis	IHC	No	[Scartozzi et al. 2009]
Her3	Melanoma	Correlation with increased cell proliferation, tumor progression and  reduced survival in melanoma patients	IHC	No	[Reschke et al. 2008]
ING3	Melanoma	Reduced nuclear expression associated with poor disease-specific survival  in melanoma patients	IHC	No	[Wang et al. 2007]
ING4	Melanoma	Reduced levels of ING4 in melanoma patients is associated with melanoma  thickness, ulceration and poor disease-specific survival and overall survival	IHC	No	[Li et al. 2008]
Ki-67	Bladder	Expression of Ki-67 is associated with progression and recurrence in  bladder cancer	IHC	No	[Gerdes et al. 1984]
	Breast	Expression of Ki-67 is associated with proliferation and progression in  breast cancer	IHC	No	[Dowsett et al. 2007]
K-ras	NSCLC	K-ras mutation is associated with poor prognosis in NSCLC patients	SA	Yes	[Zhu et al. 2008]
LOH at 18q	CRC	Associated with metastasis and poor prognosis in colorectal tumors	PCR	No	[Watanabe et al. 2001]
MammaPrint	Breast	A 70-gene prognostic assay used to identify breast cancer cases at the  extreme end of the spectrum of disease outcome by identifying patients  with good or very poor prognosis	Microarray	Yes	[Van ‘t Veer et al. 2002]
Mammostrat®	Breast	This standard purely prognostic test uses five antibodies with manual slide  scoring to divide cases of ER-positive, lymph node negative breast  cancer tumors treated with tamoxifen alone into low-,  moderate- or high-risk groups	IHC	Yes	[Ring et al. 2006]
MMP-2	Bladder	Expression of MMP-2 is associated with poor prognosis in bladder  cancer patients	PCR	No	[Xu et al. 2002]
MSI status	CRC	High frequency MSI colorectal tumors are associated with better  prognosis and show improved relapse-free survival	IHC	No	[Kim et al. 2007]
NCOA3	Melanoma	Increased levels in melanoma patients correspond to poor relapse-free  survival and disease-free survival	IHC	No	[Rangel et al. 2006]
Oncotype DX	Breast	A 21-gene multiplex test used for prognosis to determine 10-year disease  recurrence for ER-positive, lymph node negative breast cancers using a  continuous variable algorithm and assigning a tripartite recurrence score	qRT-PCR	Yes	[Goldstein et al. 2008;  Paik et al. 2004]
p21	Bladder	Overexpression of p21 is associated with poor prognosis	IHC	No	[Stein et al. 1998]
p53	Bladder	Overexpression of p53 is associated with poor prognosis	IHC	No	[Schmitz-Drager  et al. 2000]
	NSCLC	High expression of p53 in NSCLC patients confers worse prognosis  in untreated patients	IHC	No	[Tsao et al. 2007]
	NSCLC	TP53 mutation in NSCLC patients is associated with worse prognosis	SA	No
PR	Breast	Patients with PR-positive breast tumors have better survival than patients  with hormonal-negative tumors	IHC	Yes	[Dowsett et al. 2006]
Rb	Bladder	Overexpression of Rb is associated with poor prognosis	IHC	No	[Logothetis et al. 1992]
RRMI	NSCLC	High expression of RRMI in NSCLC patients confers better prognosis  in untreated patients	AQUA	No	[Zheng et al. 2007]
VEGF	RCC	Overexpression of VEGF is associated with poor prognosis in clear cell  renal carcinoma patients	IHC	Yes	[Oldenhuis et al. 2008]

AQUA, automated quantitative analysis; CA19-9, carbohydrate antigen 19-9; CAIX, carbonic anhydrase IX; CEA, carcinoembryonic antigen; CRC, colorectal tumor; CTC, circulating tumor cells; EGFR, epidermal growth factor receptor; ER, estrogen receptor; FDG, 18F-fluorodeoxyglucose; FISH, fluorescent in situ hybridization; GIST, gastrointestinal stromal tumor; IHC, immunohistochemistry; LOH, loss of heterozygosity; MMP-2, matrix metalloproteinase-2; MSI, microsatellite instability; NSCLC, non-small cell lung cancer; OS, overall survival; PET, Positron emission tomography; PFS, progression-free survival; PR, progesterone receptor; qRT-PCR, quantitative real time polymerase chain reaction; Rb, retinoblastoma; RCC, renal cell carcinoma; RRMI, ribonucleotide reductase messenger 1; SA, sequence analysis; VEGF, vascular endothelial growth factor.

Table 2.

Predictive biomarkers for cancer medicine.

Predictive biomarker	Type of cancer	Clinical significance	Detection	Clinical use	Reference
BRCA1	NSCLC	High expression of BRCA1 in NSCLC patients predicts resistance  to cisplatin-based chemotherapy	qRT-PCR	No	[Cobo et al. 2008]
	Breast	High expression of BRCA1 in breast cancer can predict response  to chemotherapy	IHC	Yes	[James et al. 2007]
CAIX	RCC	Expression of CAIX in renal cell carcinoma is predictive of sensitivity  of treatment with interleukin-2 therapy	IHC	No	[Tunuguntla and  Jorda, 2008]
c-KIT	GIST	GIST patients carrying the mutation on exon 11 of the c-KIT gene  benefit from imatinib and sunitinib treatment, however most  patients develop resistance to these over time	SA	Yes	[Dematteo et al. 2009]
EGFR1	NSCLC	EGFR1 mutations in patients with NSCLC are predictive for response  to either gefitinib or erlotinib treatment	IHC	Yes	[Sequist et al. 2007]
	CRC	EGFR1 gene amplification appears to be a predictive factor for response  to anti-EGFR1 antibody treatment in CRC	PCR	Yes	[Amado et al. 2008]
ER	Breast	High cellular expression of ER predicts benefit from tamoxifen-based  chemotherapy	IHC	Yes	[Paik et al. 2006;  Early Breast Cancer  Trialists’ Collaborative  Group, 1998]
ERCC1	NSCLC	High expression of ERCC1 in NSCLC patients predicts resistance  to cisplatin-based chemotherapy	IHC	No	[Olaussen et al. 2006;  Arriagada et al. 2004]
Her2/neu	Breast	Breast cancer patients with Her2/neu overexpressing tumors benefit  from treatment with trastuzumab in the metastatic as well as in the  adjuvant setting	FISH	Yes	[Mass et al. 2005]
	Gastric	Expression of Her-2/Neu in gastric cancer is predictive of patient  sensitivity towards treatment with 5-FU, doxorubicin, trastuzumab  and platinum-based chemotherapy	FISH	No	[Scartozzi et al. 2009]
K-ras	NSCLC	K-ras mutation positivity in NSCLC patients predicts lack of benefit  from adjuvant chemotherapy in early disease and resistance to  treatment with EGFR TKI in advanced disease	SA	Yes	[Mascaux et al. 2005]
	CRC	K-ras mutation positivity in stage IV CRC patients predicts considerably  less benefit from EGFR-specific antibody like cetuximab  and panitumumab	PCR	Yes	[Amado et al. 2008]
LOH at 18q	CRC	Useful in identifying patients with resected stage III colon cancer  most likely to benefit from 5-FU based adjuvant chemotherapy	PCR	No	[Watanabe et al. 2001]
MGMT	Glioblastoma	Methylation of MGMT promoter is predictive of sensitivity of glioblastoma  to temozolomide	PCR	No	[Dunn et al. 2009]
NuvoSelect	Breast	A combination of several pharmacogenomic genesets used primarily  to guide selection of therapy in breast cancer patients. This test also  provides the ER and HER2 mRNA status	Microarray	Yes	[Rouzier et al. 2005;  Ayers et al. 2004]
p53	NSCLC	High p53 expression in NSCLC patients predicts sensitivity  to cisplatin-based chemotherapy, however TP53 mutation is predicitve  of resistance to cisplatin-based chemotherapy	IHC/SA	No	[Tsao et al. 2007;  Winton et al. 2005]
PR	Breast	High cellular expression of PR predicts benefit from tamoxifen-based  chemotherapy	IHC	Yes	[Oldenhuis et al. 2008;  Dowsett et al. 2006;  Elledge et al. 2000]
Roche  AmpliChip	Breast	Low expression of CYP2D6 predicts resistance to tamoxifen-based  chemotherapy in breast cancer patients	Microarray	Yes	[Hoskins et al. 2009]
Rotterdam  Signature	Breast	A 76-gene assay used to predict recurrence in ER-positive breast cancer  patients treated with tamoxifen	Microarray	Yes	[Wang et al. 2005]
RRMI	NSCLC	High expression of RRM1 in NSCLC patients predict resistance  to cisplatin-based chemotherapy	qRT-PCR	No	[Rosell et al. 2004]
TP	GIST	Predicitve of senstivity of treatment to 5-FU- and capcetabine-based  chemotherapy in gastric cancer patients	IHC/PCR	No	[Scartozzi et al. 2009]
	CRC	Expression of TP in metastatic colorectal patients is predictive of sensitivity  of treatment to 5-FU and capcetabine based chemotherapy	IHC/qRT-PCR	No	[Meropol et al. 2006;  Metzger et al. 1998]
PTEN	Breast	PTEN mutation can result in reduced sensitivity of treatment with  trastuzumab in breast cancer patients	IHC	No	[Nagata et al. 2004]

CAIX, carbonic anhydrase IX; CRC, colorectal tumor; EGFR, epidermal growth factor receptor; ER, estrogen receptor; ERCC1, excision repair cross-complementation group 1; FISH, fluorescent in situ hybridization; GIST, gastrointestinal stromal tumor; IHC, immunohistochemistry; LOH, loss of heterozygosity; MGMT, O6-methylguanine-DNA methyltransferase; NSCLC, non-small cell lung cancer; PCR, polymerase chain reaction; PR, progesterone receptor; RRMI, ribonucleotide reductase messenger 1; qRT-PCR, quantitative real-time polymerase chain reaction; RCC, renal cell carcinoma; SA, sequence analysis; TK1, tyrosine kinase inhibitor; TP, thymidine phosphorylase.

This review will summarize the current excitement in this field. It is organized around the molecular technologies used to generate DNA, epigenetic, RNA, signaling pathway, protein, and metabolic tumor biomarkers (summarized in Table 3). It will discuss the limitations and future potential of these biomarkers, and how they may be productively combined with clinicopathological data. This review will also summarize the challenges of tissue collection, the use of circulating tumor cells and metabolic imaging, and the expanding role of biomarkers in drug development. Finally, it will sound a cautionary note about the need to develop a stronger evidence base including robust clinical validation prior to commercializing predictive and prognostic markers for cancer medicine.

Table 3.

The range of tumor biomarker technologies used to generate cancer predictors and prognosticators, with references to examples discussed in this review.

Measurement target	Technology	Reference
Single genes	FISH	[Lebeau et al. 2001]
	CISH and SISH	[Penault-Llorca et al. 2009]
Sets of genes	FISH	[Davis et al. 2007]
	DNA sequencing	[Ley et al. 2008]
	DNA SNP analysis	[Tenesa and Dunlop, 2009;  Easton et al. 2007]
Methylated gene promoters	Methylation-specific  quantitative PCR	[Huang et al. 2009]
Pairs of mRNAs	Quantitative PCR	[Ma et al. 2004]
Sets of mRNAs	Quantitative PCR	[Paik et al. 2004 #2974]
	Microarrays	[Liu et al. 2007; Sotiriou et al. 2006;  Wang et al. 2005; Van ‘t Veer et al. 2002]
	RNA-seq	[Shah et al. 2009]
	Microelectrodes	[Fang et al. 2009]
Tumor classification inferred from mRNAs	Microarrays	[Sorlie et al. 2001; Perou et al. 2000]
Tumor grade inferred from mRNAs	Microarrays	[Sotiriou et al. 2006a]
Tumor signaling pathways  inferred from mRNAs	Microarrays	[Bild et al. 2006a; West et al. 2001]
Exon usage	Exon-specific micorarrays	[Andre et al. 2009]
Sets of miRNAs and ncRNAs	Quantitative PCR	[Zhu et al. 2009]
Single proteins	Immunohistochemistry	[Pertschuk et al. 1979]
Sets of proteins	Immunohistochemistry	[Ring et al. 2006]
Metabolites	Biochemical assays	[Sreekumar et al. 2009]
Tumor functional pathways	Image-based probes such as PET	[Contractor and Aboagye, 2009]
Clinical data	Decision support tools	[Ravdin et al. 2001]
Clinical and mRNA data	Integrative models	[Acharya et al. 2008; Cardoso et al. 2008]

CISH, chromogenic in situ hybridization; FISH, fluorescent in situ hybridization; miRNAs, micro-RNAs; ncRNAs, noncoding RNAs; PCR, polymerase chain reaction; PET, positron-emission tomography; SISH, silver-enhanced in situ hybridization; SNP, single nucleotide polymorphisms.

The growing role of DNA sequencing

DNA sequencing technologies will play a growing role in the implementation of cancer predictors and prognosticators. At present, DNA sequencing is more common in research than in the clinic, however it is likely to become more important for patient care over the next 5 years. While high costs may prevent whole-genome sequence analyses from becoming routine for all cancer patients, ‘exome’ sequencing (targeted sequencing of protein-coding regions [Ng et al. 2009]) may become a cost-effective way to provide access to the entire transcribed genome of individuals and their tumors. To date, only a small number of whole-genome tumor sequencing projects have been undertaken. For example, Ley and colleagues sequenced the tumor genome of a patient with acute myeloid leukemia, and identified more than 500 nonsynonymous somatic mutations, including 10 mutations within protein-coding genes [Ley et al. 2008]. Recently, the International Cancer Genome Consortium (ICGC) announced a project to sequence the full genomes of 25,000 tumor samples from 50 different cancer types, with matching samples from healthy individuals (http://www.icgc.org/), to accelerate research into the causes and control of cancer.

Recently, several technological advances have increased the utility of genotyping in cancer. ‘Deep’ DNA sequencing, also known as ‘next generation’ or ‘massively parallel’ sequencing, is rapidly taking over from traditional ‘Sanger’ DNA sequencing. Essentially, in next-generation sequencing, an entire genome can be fragmented and then sequenced in parallel. As the technology can only read short lengths of DNA (<250 bp), the genome must be covered several times to ensure reliability. These advances have led to a surge in sequencing-based studies, and the full genomes of over 50 people, including related individuals with rare disorders, are now either published or in preparation [Hayden, 2009a]. Three companies are currently using different technologies to provide these next generation DNA sequences: the 454 system from Life Sciences (Branford, Connecticut, USA), Solexa from Illumina (San Diego, California, USA), and SOLiD from Applied Biosystems (Foster City, California, USA). However, even faster and less expensive ‘single-molecule’ sequencing technologies are now arriving which will further revolutionize the molecular analysis of neoplasia [e.g. Bowers et al. 2009].

Tumor genomic heterogeneity is not well captured by simply sequencing tumor tissue, since the profiles of small regions within tumors, which contain additional mutations and may eventually overgrow the tumor, are diluted. However, new methods that independently sequence multiple representatives of a given DNA fragment [Thomas et al. 2006] allow detection of DNA sequence variations present in a minority of cells within heterogeneous tumors. These advances in profiling tumor heterogeneity are complimented by methods such as COLD Polymerase Chain Reaction (PCR) [Li and markrigiorgos, 2009]that selectively amplify minority alleles from mixtures of wild-type and mutation-containing sequences [Li et al. 2009], and by advances in RNA sequencing (RNA-seq) methods described below.

Defined predictive and prognostic genetic variants

The identification of DNA variants such as single nucleotide polymorphisms (SNPs) is becoming increasingly routine, including limited genotyping of tumor DNA and screening of somatic (non-tumor) DNA for mutations that predispose to cancer or alter treatment response. Some USA hospitals now undertake broad genetic testing of almost all patients with cancer. For example, Massachusetts General Hospital in Boston is screening for 110 mutations in 13 cancer-related genes [Hayden, 2009b]. Genetic variants are also used to predict toxicity, for example polymorphisms in the UGT1A1 gene can be used to predict irinotecan toxicity [Palomaki et al. 2009]. Tumor genotyping is also offered privately; the Californian company CollabRx offers a service that analyses 15,000 genes (http://collabrx.com/).

Genome-wide association (GWA) studies are feeding into the clinical use of DNA variants. For example, GWA studies have identified cancer-causing mutations in breast [Easton et al. 2007] and colon [Tenesa and Dunlop, 2009] tumors. Somatic genetic screens are also identifying predictors of radiation sensitivity [Barnett et al. 2009] and the pharmacodynamics of anticancer drugs [Sawyers, 2008]. This research has been made possible by technological developments that have expanded the throughput and reduced the cost of SNP-based research, and allowed SNP-based techniques to identify DNA copy number variation. Nevertheless, SNP-based research to identify oncogenic DNA abnormalities remains a significant challenge, due to the difficulty of separating these cancer-causing abnormalities from genetic and epigenetic ‘noise’ [Chin and Gray, 2008]. One of the limitations of GWA studies is that we are capturing only common SNPs, which by themselves may only contribute a small amount of risk to developing cancer. To identify those rare SNPs that contribute a large amount of risk to developing cancer for a small percentage of patients, more cost-effective technologies will be required [Shelling, 2009].

Fluorescent in situ hybridization (FISH)

FISH provides an alternative way to identify predictive or prognostically important genetic variants in cancer. The most common use of FISH has been to detect copy number changes that predict treatment response, for example changes to the c-erb B2/neu (ERBB2) gene that encodes the HER2 protein [Lebeau et al. 2001; Ross, 2009b]. However, the use of FISH is now expanding. A FISH assay based on the copy numbers of three genes has been successfully used to predict prognosis in breast cancer. This test uses a proprietary algorithm to integrate the information from the three genes and predict recurrence rates. Interestingly, different three-gene sets are required for hormone-receptor positive and hormone-receptor negative tumors [Davis et al. 2007]. This is commercialized as the eXagen BC assay. Improvements have been made to FISH in the form of chromogenic in situ hybridization (CISH) and silver-enhanced in situ hybridization (SISH). These techniques use peroxidase enzyme-labelled probes whose signals do not decay over time and allow the specimen to be viewed using bright-field microscopy. CISH and SISH have been used to assess ERBB2 gene status [Penault-Llorca et al. 2009].

Gene promoter modification profiles

Epigenetics can be defined as the field of inheritable changes in gene expression that are not caused by alterations in DNA sequence. A key mechanism of epigenetics is the altered methylation of tumor-suppressor genes and of the genes encoding some micro-RNAs (miRNAs), as well as altered methylation and acetylation of the histones associated with these genes [reviewed in Esteller, 2008]. Epigenetic alterations early in tumor development may provide important predictive and prognostic tools, especially in situations where epigenetic therapies such as HDAC inhibitors are being used to reactivate gene expression [Kanai, 2007]. For example, in breast cancer, several genes associated with tumorigenesis are frequently methylated, including: RASSF1A, HOXA5, TWIST1, CCND2, p16, BRCA1, as well as genes encoding the estrogen receptor (ESR1) and the progesterone receptor (PGR) [Dworkin et al. 2009]. Epigenetic changes in several other tumor types may also provide prognostic and predictive profiles, including: ovarian cancer [Huang et al. 2009], prostate cancer [Bastian et al. 2004], glioblastoma [Nagarajan and Costello, 2009] and cutaneous tumors [Li et al. 2009]. This is an exciting field that is likely to grow in clinical importance. It may even be possible to identify promoter methylation of tumour cell DNA in body fluids such as sputum for early cancer detection [Belinsky, 2004].

An overview of predictive and prognostic mRNA profiles

A large number of profiles based on the abundance of mRNAs have been put forward as predictors of prognosis or treatment response in cancer. The simplest of these profiles are based on ratios of two mRNAs, such as HOXB13 versus IL17BR. These ratios appear to be more effective than the single mRNAs for predicting prognosis or therapeutic response [Ma et al. 2004]. The HOXB13 / IL17BR ratio has been commercialized as the Theros H/I assay. Whole-transcriptome microarray analysis of tumors has allowed predictive and prognostic molecular profiles to be generated that utilize tens or hundreds of genes. These large mRNA profiles have been particularly successful in breast cancer [Andre and Pusztai, 2006; Chang et al. 2005; Weigelt et al. 2005; Ayers et al. 2004; Paik et al. 2004, 2006; Van De Vijver et al. 2002; Van ‘t Veer et al. 2002] and colorectal cancer [Anjomshoaa et al. 2008; Garman et al. 2008; Lin et al. 2007a, 2007b; Alvarado et al. 2006; Gordon et al. 2006; Vallbohmer et al. 2006; Ahmed, 2005; Shih et al. 2005; Arango et al. 2004; Li et al. 2004; Huerta et al. 2003; Mariadason et al. 2003; Galon et al. 2002; Hegde et al. 2001]. However, they have also been applied to several other tumor types [Mengual et al. 2009; Wuttig et al. 2009; Bloomston et al. 2007; Jaeger et al. 2007; Mandruzzato et al. 2006; Winnepenninckx et al. 2006] including tumors of unknown origin [Van Laar et al. 2009]. Profiles have also been generated that can predict the progression of several different tumor types [Basil et al. 2006].

Examples of mRNA profiles used to predict prognosis and treatment response in breast cancer

As discussed in the introduction, there are a large and rapidly growing number of publications about predictive and prognostic profiles in cancer. A high proportion of these publications focuses on mRNA-based-profiles. It is beyond the scope of this review to describe all of these mRNA-based profiles, however, an illustrative selection is described below. The 21-mRNA Oncotype DX quantitative PCR assay was generated from a 25,000-gene microarray study and is used to calculate recurrence scores in breast cancer [Goldstein et al. 2008; Paik et al. 2004]. This assay estimates the 10-year recurrence risk in patients with ER^+ve, lymph node^−ve breast tumors, and has been especially useful for identifying patients who have a low risk of recurrence. The mRNAs selected for this assay place most weight on the ER and HER2 signaling pathways. The ability of Oncotype DX to guide treatment selection is formally being evaluated in the TAILORx clinical trial [Zujewski and Kamin, 2008].

The 70-mRNA MammaPrint profile [Van ‘t Veer et al. 2002] is, unlike Oncotype DX, a microarray-based test. Although this test was originally criticized for the inclusion of some patients in both training and test groups, it has since been clinically validated to a high standard and has US Food and Drug Administration (FDA) approval. MammaPrint is currently used in the prospective MINDACT clinical trial, in which its molecular-based predictions are combined with predictions based on clinco-pathological knowledge through the Adjuvant! Online decision support tool (see below) [Cardoso et al. 2008].

The Rotterdam profile [Wang et al. 2005] uses a 76-mRNA signature weighted towards proliferation-associated signaling pathways, and has no intersection with mRNAs used in either the Oncotype DX or MammaPrint tests. Another example is the 186-mRNA invasiveness prognostic signature based on the mRNA expression characteristics of CD44^high/CD24^low tumor cells [Liu et al. 2007]. The development of these and other predictive and prognostic mRNA profiles in breast cancer have been summarized in a number of excellent reviews [e.g. Bukhari and Akhtar, 2009; Geyer and Reis-Filho, 2009; Kim et al. 2009; Ross, 2009a; Sotiriou and Pusztai, 2009; Rakha et al. 2008; Ross et al. 2008].

mRNA profiles that provide classification and grading information

In breast cancer, sets of genes identified through mRNA profiling have been used to generate a ‘molecular portrait’ of tumors and have provided an informative classification into five oncogenic subtypes [Kapp et al. 2006; Sorlie et al. 2001; Perou et al. 2000]. For example, patients with ‘basal-like’ and ‘HER2’ breast tumor subtypes are more likely to have pathologic complete response after neoadjuvant multi-agent therapy [Rouzier et al. 2005]. A 50-mRNA classifier to determine these subtypes is marketed as the Breast BioClassifier (http://www.bioclassifier.com/). A set of 97 mRNAs identified through transcriptome profiling have also allowed the molecular grading of breast cancer [Sotiriou et al. 2006], which is currently being developed commercially as the MapQuant D× Genomic Grade assay.

Signaling pathway information derived from mRNA profiles

It is widely accepted that cell fate and function are regulated at the level of signaling networks and pathways rather than at the level of individual molecules [Thiagalingam, 2006]. For this reason, clinically useful molecular profiles often utilize information about the activity of entire signaling pathways [Bild et al. 2006a]. Examples include identification of BRAF-mutant tumors that will respond to inhibitors of MEK (an enzyme in a signaling pathway downstream of BRAF) [Solit et al. 2006], and the identification of PTEN deficient tumors likely to be resistant to trastuzumab (PTEN regulates the PI3 kinase signaling that occurs downstream of the epidermal growth factor receptor [EGFR] that is targeted by trastuzumab) [Nagata et al. 2004].

In breast cancer, mRNA profiles that are surrogate markers for the activity of a number of key signaling pathways have been developed [Chang et al. 2009; West et al. 2001]. Some of these mRNA profiles appear to feed directly into activities of tumorigenesis, such as E2F-associated pathways feeding into tumor cell proliferation [Hallstrom et al. 2008]. Interestingly, the activity of several specific oncogenic pathways in breast cancer appears to be associated with patient age [Anders et al. 2008]. Information about deregulation of these pathways has been useful both in breast tumor classification and in predicting clinical sensitivity to therapeutic agents [Salter et al. 2008; Bild et al. 2006b; Huang et al. 2003]. The fact that these pathway-level molecular profiles are based on known biology is very attractive.

Technologies to generate predictive or prognostic profiles from mRNA

Profiles involving small numbers of mRNAs can utilize quantitative PCR, which has high sensitivity and dynamic range and has become a mainstay of molecular profiling in cancer. PCR is used for several commercialized mRNA signatures, for example [Ma et al. 2004; Paik et al. 2004]. However, like all molecular techniques, quantitative PCR does have its limitations, as reviewed by Bustin and colleagues, such as dependence on template quality [Bustin and Nolan, 2004]. Microarrays can capture parallel information about many more mRNAs than RT-PCR—up to whole transcriptome levels. Although microarrays are expensive, susceptible to false-positive error and require complex normalization and interpretation, they have been highly productive research tools and are also used in several current prognostic and predictive tests. The clinical value of mRNA microarrays is likely to continue to improve due to concerted efforts to improve data quality [Shi et al. 2006], although they are likely to be replaced in the next 10 years by the emerging technologies described below.

Profiles based on simple mRNA quantification only capture part of the available information. More than 60% of human mRNAs undergo alternative splicing [Clark et al. 2007] yielding hundreds of thousands of mRNA transcript variants that may have distinct functions. This alternative splicing appears to be especially relevant to cancer [Venables, 2006; Faustino and Cooper, 2003]. Recently, microarrays that can interrogate over one million human exons have become avilable, with approximately four measurements per exon and on average 40 measurements per gene [Abdueva et al. 2007]. Alternative mRNA splicing data generated using this technology may be useful for predicting tumor prognosis and treatment response. For example, exon-level microarray analysis has revealed mRNAs that are differentially spliced in breast [Andre et al. 2009], colorectal [Gardina et al. 2006] and lung [Xi et al. 2008] tumors, which may contribute to tumor progression.

Next-generation DNA sequencing technology has expanded into the realm of RNA, in which case it is known as ‘transcriptome sequencing’ or ‘RNA-seq’. RNA-seq allows simultaneous analysis of all RNA molecules within a cancer cell, including alternative splice variants, mRNAs, noncoding RNAs (ncRNAs) and miRNAs [Wang et al. 2009b]. It is expected that in the future, RNA-seq will revolutionalize the analysis of RNAs in tumors, leading to several editorials suggesting that this is ‘the beginning of the end for microarrays’ [Shendure, 2008]. Due to issues with cost and analysis challenges, the replacement of microarrays with RNA-seq may still be some years away, and is unlikely to be complete. One recent example however, shows the remarkable ability of RNA-seq to analyze tumor RNA. Shah and colleagues sequenced the transcriptome of four biopsies from granulosal cell ovarian tumors, and identified a single gene, FOXL2, that was mutated in all tumors, but not in other ovarian tumors [Shah et al. 2009]. Furthermore, RNA-seq has even been used to characterize single tumor cells [Tang et al. 2009]. This, along with recently developed microfluidic devices capable of measuring the mRNA expression in single cells and the COLD PCR method [Li and markrigiorgos, 2009] discussed above, may represent a significant step towards quantifying tumor heterogeneity during prognostication [Toriello et al. 2008].

Nanotechnology is another area likely to revolutionize the use of predictive cancer markers in the next 10 years. For example, nucleic acid probes immobilized on arrays of nanostructured microelectrodes within integrated circuits are already a reality [Soleymani et al. 2009]. These systems have been used to successfully detect gene fusions in prostate cancer [Fang et al. 2009]. These technologies are in commercial development and some commentators suggest they will make PCR-free nucleic acid-based treatment response prediction available to primary care medical practitioners.

Technologies to select mRNAs for inclusion in tumor profiles

An advantage of predictive and prognostic profiles that utilize large numbers of mRNAs is their inherent redundancy—failure of one or more measurements may be compensated for by the other measured mRNAs. However, the optimal size for a clinically useful set of predictive mRNAs remains a matter for debate. It is likely that future advances will take integrative tumor pathology even further than the pathway-level understanding described above. As Karsdal and colleagues suggest, although our current understanding has progressed to the level of molecular pathways, further benefits are likely if we can extend our knowledge to a ‘systems-level’ understanding of the pathophysiology of specific diseases [Karsdal et al. 2009]. Further mathematical algorithm developments for selecting optimal mRNA combinations for different predictive and prognostic tasks will be important. Currently, methods used range from simple classification systems to highly sophisticated machine learning methods [Sotiriou and Piccart, 2007; Saidi et al. 2004]. The ever-expanding systems biology databases such as the Ingenuity Pathway Analysis database (http://www.ingenuity.com/) and the Kyoto Encyclopedia of Genes and Genomes (KEGG) [Kanehisa et al. 2006] may also be useful for the development of future pathway-specific molecular profiles. However, all of these methods require more formal evaluation than has been performed to date.

Limitations of mRNA-based predictive and prognostic molecular profiles

Despite the significant body of literature describing predictive or prognostic mRNA profiles for cancer, only a small number are used in current oncology practice; examples are shown in Tables 1 and 2. In fact, as few as 3% of published studies describing potential clinical applications in genomic medicine have progressed to a formal assessment of clinical utility [Khoury et al. 2007]. It seems that mRNA-based molecular prognostic or predictive tests for cancer have not yet achieved their full potential due to a variety of limitations including: bias [Ransohoff, 2005], poor ability to generalize [Reid et al. 2005], instability of gene lists [Koscielny, 2008], inadequate sample sizes [Sorlie et al. 2006], inadequate statistical analysis [Faratian and Bartlett, 2008] and insufficient numbers of studies testing their therapeutic utility [Tavassoli, 2009; Simon et al. 2003]. Other limitations include predominantly retrospective and observational study designs [Tinker et al. 2006], poor representation of small early-stage tumors, and in breast cancer specifically, a limited predictive spectrum beyond ER^+ve tumors [Geyer and Reis-Filho, 2009]. An additional criticism is that many mRNA-based predictors have grown out of academic projects rather than being designed ‘from the ground up’ to support key therapeutic decisions [Simon, 2005]. Kim and colleagues suggest that “… clinically useful prognostic and predictive markers are those developed with a specific clinical context in mind and tested and validated within that clinical context” [Kim et al. 2009]. Views about the use of molecular profiling in oncology have become polarized. As one commentator has observed, “The field of microarray expression profiling is burdened with both unrealistic hype and excessive skepticism” [Simon, 2005].

MicroRNAs and noncoding RNAs

The analysis of RNA has recently expanded into the field of ncRNA. As their name implies, ncRNA molecules do not encode protein products, nevertheless they are important regulators of gene expression, and are able to directly affect the phenotypes of cells. There are several types of ncRNAs, of which miRNAs are the most widely studied and characterized. miRNAs have become a hot topic for researchers in every field since their discovery in Caenorhabditis elegans in 1993 [Lee et al. 1993]. A link between them and cancer was sealed with the report of loss of two miRNA genes at 13q14 in chronic lymphocytic leukemia [Calin et al. 2002]. Since this report, the expression of many individual miRNAs has been associated with patient survival, drug treatment response and tumor metastases in a number of different cancers [Sotiropoulou et al. 2009; Blenkiron and Miska, 2007]. For example, in breast cancer, specific miRNAs are associated with the five mRNA-derived tumor subtypes [Blenkiron and Miska, 2007]. Particular miRNAs may also be signatures of early carcinogenesis [Kalscheuer et al. 2008]. miRNAs are very stable in tumors, even in archival specimens, and in body fluids such as blood and serum [Chen et al. 2008; Mitchell et al. 2008]. Specific miRNAs have been sequenced from the circulating blood of patients with lung cancer and colorectal cancer [Chen et al. 2008], and for example, miR-155 may be differentially expressed in the serum of women with PR^−ve compared to women with PR^+ve breast cancer [Zhu et al. 2009].

In contrast to miRNAs, much less is known about the other types of ncRNAs, with novel ncRNAs continuing to be discovered [Guffanti et al. 2009]. Many ncRNAs also appear to be differentially expressed in specific tumor types (reviewed in [Mallardo et al. 2008]), and with the advent of RNA-seq, the possibility to discover as yet unknown but potentially important ncRNAs is high. All of this raises the exciting prospect that ncRNAs may in the future provide clinically-useful biomarkers for cancer prognostication [Bartels and Tsongalis, 2009].

Proteins

Immunohistochemistry (IHC) of tumor tissue has been a cornerstone of protein-based tumor marker work for several decades [Pertschuk et al. 1979]. A full description of IHC in cancer is outside the scope of this review; however, we will briefly summarize some topical points. IHC has the advantages of providing morphological information about protein expression and is of comparatively low cost, but has the disadvantages of sensitivity to tissue processing, scoring variability [Ross et al. 2007] and a limited availability of validated markers. One of the most interesting developments is the use of multiple antibodies to generate synergistic data for several proteins. For example, IHC data for the proteins p53, NDRG1, CEACAM5, SLC7A5, and HTF9C appears to be superior to data from currently-used single markers for predicting outcome in ER^+ve breast cancer [Ring et al. 2006]; this is now commercialized as the Mammostrat assay.

The area of proteomics has long promised biomarkers for the early detection of cancer. For example, phosphospecific antibodies can be used to identify specific kinase activity [Cloughesy et al. 2008], and a large-scale survey of tyrosine kinase activity in lung cancer has identified several novel molecular features of lung cancer that are invisible at the DNA or RNA level [Rikova et al. 2007]. Low molecular weight circulating blood proteins—known as the ‘peptidome’—may also provide potential prognostic markers for early tumors [Petricoin et al. 2006]. Microvesicles shed by tumors carry molecular signatures such as variant EGFR mRNAs, and may provide additional sources of predictive and prognostic information [Skog et al. 2008].

Unfortunately, several research groups have reported potential protein/proteome-based biomarkers, only to find later that they have not reached clinical utility after greater levels of scrutiny. Recently, the prostate-specific antigen (PSA) has been shown in two large clinical trials to be largely ineffective for early detection of prostate cancer, and to lead to over diagnosis and overtreatment [Barry, 2009]. Robust clinical evaluation is likely to be needed before the potential of proteomic biomarkers is fully realized.

Metabolite profiling

Metabolomics, which quantifies the metabolite content of cells or tissues, is also potentially useful for predicting treatment response, since metabolites represent the destination or endpoint of many molecular pathways. The recent discovery that sarcosine, a derivative of glycine, is elevated in the urine of men with metastatic prostate cancer [Sreekumar et al. 2009] raises the hope that it will be of benefit for the detection of aggressive cancer. Developments in imaging modalities (see below) that allow detection of tumor metabolites are likely to accelerate this field [Brindle, 2008].

Molecular imaging

Tumor biology and pharmacology is increasingly amenable to in vivo evaluation using image-based molecular probes. Positron-emission tomography (PET) offers some of the most promising examples in this area. The most commonly used PET tracer, ¹⁸F-fluorodeoxyglucose (FDG), is a marker of glucose transport in gastrointestinal stromal tumors (GIST) [Contractor and Aboagye, 2009]. There are many other image-based biomarkers relevant to the biology of living tumor cells, at various stages of laboratory and clinical development. These are showing promise in several applications, including the detection of occult tumors. Although a full discussion is beyond the scope of this paper, this area is reviewed in several recent papers [Josephs et al. 2009; Thakur, 2009; Hargreaves, 2008].

Decision support tools

As part of clinical acumen, physicians have traditionally integrated clinical and pathological data of several different types to make treatment decisions. However, as the amount of information available grows, there comes a stage when there is too much information to manually integrate reliably; leading to the need for formal mathematical or computational models of tumors that incorporate this information [Lazebnik, 2002]. For this reason, clinicopathological data have been integrated using various types of mathematical algorithms to provide decision support tools for clinicians [Abbott and Michor, 2006]. Decision support models are available for several tumor types, for example: urological tumors [Donovan et al. 2009; Abbod et al. 2007], HER2^+ve breast tumors [Lisboa et al. 2007], for decisions about adjuvant therapy (Adjuvant! Online, http://adjuvantonline.com/index.jsp) [Ravdin et al. 2001], and for selecting optimal strategies to manage colorectal liver metastases [Poston et al. 2005]. One group has used the term ‘systems pathology’ to describe their method of using pathological data with clinical data and limited molecular data for prognostication after prostate cancer surgery [Cordon-Cardo et al. 2007].

Combining clinical and molecular cancer predictors

A principle of the field of systems biology is to combine multiple data types in order to reduce false-positive errors [Ge et al. 2003]. It is attractive to apply this principle to prognostic or predictive profiles in cancer, where different types of molecular and/or clinical data may be synergistic, and where separately-generated molecular profiles may concord to add weight to predictions [Fan et al. 2006]. For example, in cases where quantitative PCR reports a low average abundance of a predictive mRNA in a tumor, it may be useful to know from IHC that the protein encoded by that mRNA is expressed at high levels, but only by cells in a small dysplastic region of an otherwise low-grade tumor. In another example, molecular grade signatures may assist pathologists when grading tumors that have not been optimally fixed [Sotiriou et al. 2006].

Shedden and colleagues have shown that molecular tumor classification can be made accurately using fewer mRNA measurements if information about pathology and tissue ontogeny is also included [Shedden et al. 2003]. Pittman and colleagues have combined pathway-level summaries of mRNA abundance with traditional clinical risk factors to predict breast cancer recurrence [Pittman et al. 2004]. Salter and colleagues have integrated ER and HER2 status, mRNA profiles and miRNA profiles for the NCI-60 cell lines to predict drug sensitivity [Salter et al. 2008]. Acharya and colleagues have incorporated pathway profiles from mRNA expression data with clinical risk stratification to derive prognostic ‘clinicogenomic clusters’ [Acharya et al. 2008]. In the MINDACT breast cancer clinical trial, mRNA profiles from the MammaPrint microarray assay are used alongside clincopathological knowledge contained in Adjuvant! Online, using a relatively simple voting system [Cardoso et al. 2008].

The problem of data volume

As tumor molecular profiling becomes more routine in clinical practice, the large amounts of data associated with each patient needs to be annotated, individual items of information appropriately linked to the other items of information, and this information stored in a secure but easily retrievable format [Wang et al. 2009a]. Layered on top of these challenges is the frequent involvement of patients in clinical trials, and the need for tumor molecular profiles to be available for research and regulatory purposes, with access restriction based on patient consent and ethical approval providing a further layer of complexity. In a 2003 report, the National Institutes of Health highlighted the balance between ethical use/privacy and easy access to “make data as widely and freely available as possible, while safeguarding the privacy of participants, and protecting confidential and proprietary data” [NIH, 2003]. Sharing data as freely as possible within ethical, commercial and regulatory constraints drives development of new hypotheses and fosters cross-validation, as exemplified by archives of gene expression data such as the Gene Expression Omnibus (GEO) [Edgar, 2002] and ArrayExpress [Parkinson et al. 2007]. Several bioinformatic initiatives such as the US National Cancer Institute’s cancer Biomedical Informatics Grid (caBIG) aim to link pathological, molecular, clinical and treatment response data. In the future these large initiatives, rather than single-center research studies, are likely to become the dominant platform for generating and assessing predictors and prognosticators in cancer medicine [McCormick, 2009; Hanauer et al. 2007]. A current research challenge is how to integrate diverse types of clinical and molecular information from thousands of patients. An approach that appears to have promise is probabilistic Bayesian modeling, which has been used successfully to integrate microarray data from multiple studies [Sykacek et al. 2007] and to integrate clinical and molecular data in breast cancer [Gevaert et al. 2006].

The challenge of accessing tumor tissue

The uptake of molecular predictive tests in solid tumors has been slower than in leukemia, possibly due to the more difficult access to solid tumor cells [Sawyers, 2008]. In trials involving solid tumors, often the only source of tissue is from a single diagnostic biopsy or after resection. Although not always practical, studies involving re-biopsies have been dramatically successful—for example the identification of lung cancer patients who have acquired EGFR kinase domain mutations during gefitinib treatment [Pao et al. 2005]. Studies comparing molecular profiles derived from biopsies to those from resected tumors have proved interesting—it appears that although profiles from biopsies generally overlap with those from resected tumors, they differ particularly in those molecules expressed by the tumor stroma, which may be sampled differently by each method [Symmans et al. 2003]. It is encouraging that even the small amounts of tumor tissue obtained by fine needle aspiration appear capable of predicting treatment response [Ayers et al. 2004]. This is also the attraction of noninvasive image-based testing paradigms. Due to improved diagnostic and screening procedures, some tumors are now being diagnosed as very small lesions. Many of the tumor prognostic profiles in current use originated in studies of large late-stage tumors, and therefore, even if they are clinically effective, they may not be optimal for predicting the therapeutic response of the small early-stage tumors identified in screening programs. An additional complication, even in small tumors, is tumor heterogeneity [Fey and Tobler, 1996]. To overcome this, careful selection of multiple representative tumor regions may be required [Ross et al. 2007; Paik et al. 2005; Carey et al. 1990].

Due to nucleic acid degradation and cross-linking, DNA or RNA extracted from formalin-fixed and paraffin-embedded (FFPE) tumor tissue tends to be of lower quality than DNA or RNA extracted from unfixed tissue [Paik et al. 2005]. Those assays that can analyze predictive or prognostic profiles from FFPE tumor tissue (such as Oncotype DX) may have a practical advantage over those that cannot (such as the current version of MammaPrint). There has been a gradual development of more effective protocols for recovery of DNA and RNA from FFPE tissue, and some studies report that mRNA expression profiles from FFPE tissues and fresh frozen tissues are similar [Linton et al. 2009; Furusato et al. 2007]. However, progress is still needed towards better recovery of DNA and RNA from FFPE tissue and towards better fixation processes. These developments will not only benefit predictive clinical assays, but will also benefit cancer research by unlocking a treasure trove of FFPE archival material.

Circulating tumor cells

The identification of circulating tumor cells in peripheral blood has been recognized for many years as a promising diagnostic tool for the early detection of primary tumors and metastatic deposits [Mostert et al. 2009; Sleijfer et al. 2007]. In addition, a sufficiently sensitive assay based on circulating tumor cells could be valuable for early detection of recurrent disease. Circulating tumor cells have been found in a wide range of cancers, including breast [Ross and Slodkowska, 2009], colorectal cancer [Sergeant et al. 2008] and melanoma [Nezos et al. 2009]. Reliable detection of circulating tumor cells is difficult, since they form only a very small proportion of the heterogeneous range of cells present in the peripheral blood. However, recent studies have demonstrated clinical utility. For example, the CellSearch System (Veridex, North Raritan, New Jersey, USA) has FDA approval and has been used in several clinical settings including monitoring treatment response in prostate, colon and breast cancer [Nakamura et al. 2009].

Regulation, validation and cost effectiveness

Several predictive and prognostic tests have now been commercialized for decision-making in oncology [e.g. Ross et al. 2008; Weigelt et al. 2005; Paik et al. 2004; Van ‘t Veer et al. 2002]. The experience with cetuximab provides an example of the commercial value inherent in such testing. Cetuximab has been developed as a treatment for advanced colorectal cancer by the firms ImClone, Bristol-Myers Squibb and Merck. It was evaluated in randomized trials involving several hundreds of patients [Bokemeyer et al. 2009; Van Cutsem et al., 2009]. During the course of these trials, Lièvre and colleagues published a study of 30 patients that found cetuximab had no effect in tumors carrying mutant forms of the proto-oncogene K-ras [Lievre et al. 2006]. This finding was confirmed in the larger studies, and as a consequence the market for cetuximab shrank by around 35%, the frequency of K-ras mutant tumors. This finding clearly means that anyone contemplating purchase of cetuximab to treat bowel cancer will place a high value on K-ras mutational analysis. Such predictive tests will increasingly come to represent marketable intellectual property.

The regulation of predictive and prognostic molecular tests for cancer is complex. In the USA, there is a dual path system where the majority of molecular prognostic tests (which are developed in-house by testing laboratories as ‘home brew’ tests) are not examined by the FDA, provided technical aspects of the test are certified. In contrast, those tests marketed as ‘kits’ must be fully reviewed by the FDA as medical devices [Schmidt, 2008]. Several commentators have argued that this produces a situation where many molecular tests are not supported by adequate evidence before they are marketed [Katsanis et al. 2008]. For example, a recent study of HER2 testing in breast cancer patients to predict trastuzumab response suggested that there may be concerning variations in practice and a significant incidence of inadequate documentation, as well as potentially inaccurate results [Phillips et al. 2009]. However, this view is balanced by the fact that verifying the analytical and clinical validity of cancer biomarkers is a complicated process [Schmidt, 2008; Gutman and Kessler, 2006], and that regulatory processes can greatly slow the translation of important medical advances into patient care, as well as greatly increase costs [Rogowski et al. 2009; Pitts, 2008]. It is also interesting to consider the ethical issues of developing predictive tests that, due to cost, may not be available to the majority of the world’s population, and may not even be cost effective for affluent populations [Ioannidis, 2009].

Another issue is that the calibration of a test will influence its cost effectiveness. To the extent that a test is amenable to receiver-operator analysis (analysis of the balance between sensitivity and specificity), designation of the threshold for positive versus negative results is important. Lenient thresholds will accept lower likelihoods as predicting positive benefit from a treatment, and tend to increase access. Stricter thresholds will demand a higher likelihood before predicting positive benefit, and tend to reduce access [Laking et al. 2006 #126]. Inevitably, proponents of treatments, such as the pharmaceutical industry, will have an interest in the design of tests that predict benefit from their agent. Tests designed for use in wealthier settings may not be well-calibrated for poorer settings, in that they may support less-cost-effective allocations of treatment.

Overall, it appears that the development of appropriate regulatory structures, a robust evidence base [Faratian and Bartlett, 2008] and physician preparedness for molecular prognostic tests in cancer have lagged behind the development of the tests themselves [Avard and Knopper, 2009; GenomeWeb, 2009; Scheuner et al. 2008]. It is encouraging to see the emergence of large clinical trials for some breast cancer predictive tests; the Oncotype DX test (TAILORx, [Zujewski and Kamin, 2008]) and the MammaPrint test (MINDACT, [Cardoso et al. 2008]). However, the design of these trials may not allow direct conclusions to be drawn about the clinical usefulness of these tests [Koscielny, 2008]. As more cancer biomarkers are commercialized, there needs to be increasing focus on providing an evidence base for improved clinical outcome using prospective randomized clinical trials, or if possible, using clinical trials in which randomization for specific treatment modalities is based on a biomarker. In addition, to allow for the optimal use of these assays, a wide range of clinical settings (such as locally advanced or metastatic disease) needs to be evaluated. To facilitate optimally cost-effective use of such tests, the user needs to retain the freedom to define thresholds for positive versus negative results. The cost of providing this necessary level of reassurance to oncologists may well limit the number of tests that can be developed in the future.

Prognostic and predictive molecular tests as part of new drug development

Aside from the issues discussed immediately above, cancer biomarkers have the potential to play a valuable role in future drug development, as part of the FDA’s Critical Path Initiative [Karsdal et al. 2009]. This is an enormous area, and is largely outside the scope of this review. However, a small number of examples especially relevant to cancer will be discussed. Tamoxifen activity is dependent on its conversion by the CYP2D6 enzyme to its active metabolite endoxifen, and several studies show that poor metabolizers of tamoxifen do not get as much benefit from the drug as other patients [reviewed by Hoskins et al. 2009]. The Roche Amplichip (Roche, Basel, Switzerland), which measures the genetic variation in CYP2D6 and CYP2C19, is being applied to tamoxifen treatment in an attempt to improve efficacy for individual patients. However, well-designed clinical trials will be required to confirm the Amplichip’s clinical utility in this setting [Hoskins et al. 2009]. In addition to the Amplichip, there are numerous other commercial pharmacogenetic initiatives, such as the Belgian company DNAVision (Charleroi, Belgium), that market genetic tests to predict adverse reactions to 5-fluorouracil and irinotecan in colorectal cancer (http://www.dnavision.be/).

Molecular profiling at the DNA or RNA level may also be used to enrich clinical trial participants for those patients most likely to benefit from a drug [Adjei et al. 2009]. While this may narrow the eventual market for the drug, it may also reduce the likelihood of expensive failures at late developmental phases [Gutierrez et al. 2009]. The molecular profile of likely drug responders may be known before the trial begins, or may rely on comparing tumor DNA/mRNA profiles from phase II ‘responders’ with the profiles of phase II ‘nonresponders [Maitournam and Simon, 2005]. Molecular profiles may also be used to predict pharmacodynamics. For example, in target engagement studies, instead of selecting phase II doses based on the maximum tolerated dose from a phase I dose-escalation study, phase II doses are determined by assessing the effect of a drug on the molecular profile of the drug’s targeted pathway [Sawyers, 2008].

Conclusion

It is an exciting time to be an oncologist or cancer biologist. Significant advances have been made in technologies for DNA and RNA sequencing, SNP genotyping, mRNA and ncRNA quantification, molecular pathway analysis and integrative bioinformatics. These advances have greatly expanded our understanding of the molecular basis of tumor progression and individual treatment response. Several predictive and prognostic tests based on this new knowledge have revolutionized cancer medicine. However, this field still has a long way to go before it reaches its full potential. To accelerate their benefit to patients, prognostic and predictive markers need to be developed ‘from the ground up’ with specific clinical contexts in mind, and subjected to robust clinical evaluation. Methods developed for use with small amounts of FFPE tumor material or circulating tumor cells, as well as noninvasive molecular imaging, are likely to be especially useful. We may need to pay particular attention to optimal tumor sampling and to predictive methods that formally combine clinical and molecular data into a synergistic understanding of tumor progression. Bioinformatic initiatives that combine large amounts of cancer data such as caBIG are likely to play increasingly important roles. Overarching all of these points is the need to train translational physician-scientists who have a deep appreciation of medicine, cancer biology and the business of clinical implementation. As predictors and prognosticators continue their exciting expansion into cancer medicine, we must ensure they retain a strong evidence base.

Conflict of interest statement

None declared.