Journal of the New Zealand Medical Association, 11-October-2002, Vol 115 No 1163
Cancer genomics: the chips are on the table
Kieran Holland and Michael Sullivan
Despite tremendous progress in our understanding of the genetic origins of cancer, this diverse group of diseases remains enigmatic. Patients with apparently similar tumours may experience very different outcomes. Existing methods of classifying tumour samples often tell clinicians little about how an individual patient will respond to treatment. This review discusses the role that recently developed DNA microarray (‘gene chip’) technology may play in the development of powerful new molecular classification systems for cancer.
Diagnosis: a clinical classification
Diagnosis is a form of classification that enables clinicians to make informed judgements about a patient’s prognosis and most suitable treatment options. In the broadest sense, classification is the process of grouping entities according to similarities in selected properties so that we can compare, contrast, and make predictions. The power of a classification system is limited primarily by our ability to measure properties of interest and our ability to effectively subdivide things according to those properties.
Until recently, the properties that we have used to classify cancers have been those most readily accessible: macroscopic and microscopic morphological features. These features include the organ of origin, tumour size, tissue type, cellular appearance, lymph node involvement, and evidence of distant spread. The widely used TNM (Tumour-Node-Metastasis) system, originally developed by Pierre Denoix over 50 years ago, incorporates many of these features into a standardised format that has been favoured for its utility, simplicity, and general applicability.1
However, morphological cancer classification systems are far from ideal. The subjective nature of morphological typing means that the interpretation of histological sections frequently differs between specialists.2 Within existing diagnostic groupings there are often large variations in patient outcome and, as the groupings have been refined through the addition of extra parameters, the systems are becoming increasingly complex.
Molecular ‘markers’ are now used in clinical practice to improve cancer diagnoses. In breast cancer, for example, oestrogen receptor and ERBB2 (HER-2/neu) expression is used to predict responses to tamoxifen and trastuzumab (Herceptin), respectively.3 But measuring a few isolated molecular properties of cancer cells may leave many important distinguishing features undetected, and as the number of markers increases it is becoming difficult to integrate them with standard classification systems. Recent advances may lead to significantly more powerful molecular classification systems for cancer.
The molecular view of cancer
What sets cancer cells apart from normal cells is a series of progressively acquired genetic changes that corrupt the molecular pathways designed to constrain cell growth. These genetic changes interfere with the expression of the genome, leading to changes in the protein content of the cell (proteome). For a gene to be expressed, its DNA sequence is first transcribed to create an RNA copy and then processed to produce ‘messenger’ RNA (mRNA). The ribosomal machinery of the cell translates the mRNA sequence into a protein product. The draft sequence of the human genome, published in 2001,4,5 estimated the total number of human genes to be around 32 000. However, this is a preliminary result and competing estimates have ranged between 30 000 and 120 000 genes.6
The subset of genes implicated in the development of cancer fall into two major classes: oncogenes and tumour suppressor genes. Oncogenes are normal genes that have undergone ‘gain-of-function’ mutations, resulting in pathological stimulation of cell growth. Tumour suppressor genes normally act to constrain the proliferation of abnormal cells, but in cancers may be disabled by ‘loss-of-function’ mutations. Most cancers result from multiple genetic changes involving both oncogenes and tumour suppressor genes. Genetic changes that provide the cell with a growth or survival advantage are positively selected during tumour development. Underlying these genetic events is thought to be an acquired or inherited instability in the cancer genome – the so-called mutator phenotype – that predisposes the cancer cell to the acquisition of additional mutations through errors in genome replication or repair.7
Hanahan and Weinburg have drawn together the last 25 years of cancer research to produce a framework of six functional changes necessary for a cell to become cancerous.7 These are: 1) growth independent of normal growth-stimulating signals; 2) invulnerability to the mechanism of regulated cell death (apoptosis); 3) induction of a blood supply (angiogenesis); 4) unlimited potential for replication; 5) capacity to invade other tissues; and 6) ability to grow at metastatic sites. Each cancer may acquire these functions in a slightly different way. Although mutations in oncogenes and tumour suppressor genes may be the initial trigger, the expression patterns of many more genes will be altered as a consequence, providing a unique fingerprint for each tumour. By studying global alterations of gene expression in cancer cells, it may be possible to identify the functional differences that underlie much of the variability in the behaviour of morphologically similar tumours.
While it remains difficult to study global gene expression patterns at the protein level,8 inferences about gene expression can be made from the mRNA content of the cell (transcriptome). Until recently, it was possible to study the mRNA levels of only a few genes at a time, but the advent of DNA microarray technology has made it possible to study the expression of thousands of genes at once.9
An expression microarray consists of a grid of DNA spots, each containing a sequence from a particular gene. These DNA sequences are printed onto a glass slide by robot or synthesised directly on a silicon wafer using methods similar to those for making computer chips. Figure 1 illustrates how microarrays can be used to compare the mRNA levels in a tumour tissue sample with the levels in a standardised reference sample to form an expression profile for that tumour.
Figure 1. Expression profiling at a glance
In a typical cancer expression profiling experiment, the mRNA levels in a tumour tissue sample are compared to levels in a standardised reference sample. (a) First, mRNA is extracted from the tissue samples. (b) The RNA is reverse transcribed into cDNA. The tumour cDNA is labelled with a red fluorescent dye (Cy5), whereas the reference cDNA is labelled with a green fluorescent dye (Cy3). (c) The labelled cDNA samples, containing a mix of all the expressed gene sequences, are simultaneously hybridised to a microarray slide, where they bind to the complimentary gene sequences. The amount of cDNA binding to each arrayed gene sequence is proportional to the level of expression of that gene in the tumour and reference tissues. (d) The hybridised slides are washed and then imaged by a laser scanner, which detects fluorescence in the red and green spectrums. (e) Image processing software is used to determine the ratio of the fluorescent intensity in the red and green images for each spot on the microarray. This ratio provides a measure of the relative quantity of RNA in the tumour sample compared to the reference sample for each gene sequence probed. (f) A computer-generated dendrogram clusters tumours with similar gene expression profiles together.
Expression profiles reveal so many properties of cancers that researchers have been left wondering how to make the most of this information. Can expression profiles be used to create better cancer classifications? Which genes are important? And how might this improve the management of individual patients? The first step in answering these questions has been the development of automated computational methods for identifying patterns within expression profiles.
Bioinformatics: dissecting the data
Bioinformatics is the use of mathematical and computational approaches to solving biological problems. This rapidly expanding field arose out of the need for systems to manage and analyse the large volumes of data being produced by genome sequencing and protein structure research. There are many different methods being used by bioinformaticians to identify the molecular signatures that characterise different types of cancer, but the majority fall into two groups: clustering algorithms and class prediction algorithms.10
Clustering algorithms (unsupervised learning) identify groups of tumours with similar gene expression profiles. Hierarchical clustering has been particularly popular for microarray experiments because the results can be displayed in the form of a dendrogram – a tree-like diagram where the most similar examples are grouped together on nearby branches (Figure 1, (f)). Clustering algorithms help to identify previously unrecognised similarities between tumours, but the results can vary greatly depending on which genes are included and which clustering algorithm is used.
Class prediction algorithms (supervised learning) use a representative ‘training set’ of tumour expression profiles, grouped according to some existing criteria such as diagnosis or response to treatment, to generate a tumour class predictor. The class predictor uses the gene expression pattern of a new tumour to predict which group it belongs to. The classifier is judged according to its success in classifying examples in an independent ‘testing set’. Prediction algorithms are powerful tools for identifying patterns of gene expression that correlate with known tumour properties but are less likely to reveal previously unrecognised similarities between tumours.
A new way of classifying cancer
In 1999, Golub et al showed that a class prediction algorithm could accurately distinguish between the expression profiles of acute myeloid leukaemia (AML) and acute lymphocytic leukaemia (ALL).11 Although this distinction was already possible with existing techniques, the finding validated the notion that cancers might be classified using expression profiles. Ramaswamy et al (2001) added support for this idea by showing that tissue samples from 14 different tumour types could be distinguished by their expression profiles alone.12 Now the focus of research is rapidly shifting towards clinical applications of this method.
Exposing a disease within a disease
Expression profiling has recently demonstrated what oncologists have long suspected: patients with clinically and histologically identical tumours may actually have distinct disease subtypes that were not previously detectable. Alizadeh et al (2000)13 and Shipp et al (2002)14 profiled the tumours of patients with large diffuse B-cell lymphoma and were able to identify sub-types that were associated with markedly different survival. The prognostic power of these new classifications was independent of the standard clinical staging criteria, the International Prognostic Index.15 In a similar analysis of paediatric ALL, Yeoh et al (2002) identified distinct gene expression patterns associated with known molecular subtypes and also identified a cluster of samples that may represent a new subtype of leukaemia.16
A particularly difficult problem in oncology is determining the intensity with which to treat an individual patient. One example comes from patients with low-stage breast cancer, where adjuvant chemo- and hormonal therapies are used to reduce the risk of metastases. Currently, patients are selected for treatment on the basis of clinical and histological assessment, but the selection procedure is imprecise and many patients who receive adjuvant therapy may have survived without it. Using expression profiling, van ’t Veer et al (2002) studied 117 primary breast cancers in women under the age of 55 years.17 By applying a class prediction algorithm, they were able to accurately identify the subgroup of patients most likely to benefit from adjuvant therapy and conversely the group to which it conferred no advantage. This approach provides a strategy for targeting high-risk cancer patients with more intense therapy and reducing the unnecessary treatment of low-risk patients leading to a reduction in treatment failures, side effects, and unnecessary use of costly therapies.
Assisting drug discovery
Expression profiling will also provide new targets for drug development by identifying regulatory pathways that are activated specifically in malignant cells, hopefully resulting in more powerful treatments with fewer side effects. The value of targeting cancer treatments at specific molecular abnormalities has recently been demonstrated by the success of imatinib mesylate (Glivec), a signalling pathway inhibitor that has proved highly effective in the treatment of chronic myeloid leukaemia.18
Microarrays for medicine?
Only a fraction of the cancer markers reported in the literature are ever used in the clinic. A new classification tool like expression profiling will only be adopted if it is both proven and practical for use in day-to-day clinical practice. The clinical benefits of expression profiling must first be validated in large clinical trials. Whether microarrays then become practical as a clinical tool will depend on factors such as cost, speed, and simplicity. With increasing demand for this technology, competition amongst commercial suppliers is rapidly driving the costs down. Conceivably, expression analysis could be performed within a hospital laboratory and the result reported as a simple diagnostic grouping with a prognostic index and list of predicted drug sensitivities. While it is unlikely that expression profiling will replace morphological analysis, it may provide an efficient alternative to performing multiple molecular, cytogenetic, and immunohistochemical tests.
But there are still important issues to be resolved before expression profiling is ready for clinical use. Standard protocols for tumour tissue collection will need to be adapted to cater for molecular techniques. For example, the formalin preservatives used for histology are damaging to RNA19 so a portion of fresh tumour tissue must be allocated for expression analysis, taking care not to interfere with important morphological features such as tumour margins.
The management of gene expression data remains one of the greatest challenges. The genome sequencing projects have paved the way for large-scale biological data collection but, whereas DNA sequences can naturally be represented by strings of letters, there is no equally intuitive format for gene expression profiles. Furthermore, expression profiles can vary with different array technologies and laboratory methods. In an effort to address these problems, the Microarray Gene Expression Database (MGED) group is developing standards for annotation, storage, normalisation and exchange of gene expression data.20 The widespread adoption of such standards is vital to progress in this field because at the moment it is impossible to systematically compare the results of different laboratories.
Microarray gene expression profiling promises to help define new molecular classification systems for cancer that transcend the limitations of morphological classification. Preliminary studies suggest that expression profiling will result in more reliable prognoses for cancer patients and permit individualised disease management plans. Nevertheless, early results must be confirmed in large clinical trials and technical issues remain to be solved before expression profiling is ready for clinical use.
The challenges posed by this new technology are diverse, and translation of research into better health care will require close collaboration between a wide range of specialists including clinicians, biologists, statisticians and bioinformaticians. Coupled with other developments in molecular biology, it is likely that expression profiling will have a significant impact on the practice of clinical oncology in the coming decade.
Author information: Kieran Holland, Medical Student; Michael Sullivan, Senior Lecturer in Paediatric Oncology, Department of Paediatrics, Christchurch School of Medical and Health Sciences, University of Otago, Christchurch
Acknowledgements: We thank Lana Ellis, Christine Morris, Bridget Robinson, and Logan Walker for their valuable advice and comments. Kieran is grateful for the financial aid provided by the Tassell Scholarship for Cancer Research during the course of his Bachelor of Medical Science degree.
Correspondence: Dr Mike Sullivan, Department of Paediatrics, Christchurch School of Medical and Health Sciences, P O Box 4345, Christchurch. Fax: (03) 364 0747; email: email@example.com
issue | Search journal |
Archived issues | Classifieds
| Hotline (free ads)
Subscribe | Contribute | Advertise | Contact Us | Copyright | Other Journals