Correspondence *Department of Pathology, Netherlands Cancer Institute, Plesmanlaan 121, 1066 CX Amsterdam, The Netherlands

Email m.vd.vijver@nki.nl

Introduction

The discovery of 25,000 human genes as a result of the completion of the Human Genome Project has greatly accelerated the research into genotypic–phenotypic correlations in an aim to elucidate the functional taxonomy of genes in both normal tissues and disease states. The use of microarray technology to perform gene-expression profiling is an important adjunct to our current knowledge of genetics, because microarrays allow the study of thousands of genes simultaneously in a single, standardized, and cost-efficient experiment. Gene expression is a broad term used to describe the transcription of information encoded within DNA sequences into mRNA, and the subsequent translation of the mRNA information into proteins that regulate cell function. The gene expression of all cells is a highly dynamic process that alters in response to cell requirements and environment changes, and as a consequence of disease. A brief summary of the applications of gene-expression profiling by microarray analysis and additional application of microarray platforms are described in Table 1.

Table 1 Examples of current applications of microarray technology.^a

Full tableFigures & Tables indexDownload Power Point slide (254K)

Gene-expression profiling has been used to study infectious and immunological diseases, but the predominant focus of microarray-based research has been in the study of cancer. Microarray analysis has made it possible to identify groups of genes according to their expression pattern or 'genetic signature', and this could potentially ameliorate the clinical management of a patient. Despite the vast amount of research in this field, a disparity exists between the quantity of publications and the quality of their interpretation. The quantity of articles that document the discovery of new gene profiles is as plentiful as the number of publications that scrutinize their interpretation.¹

In this article we will briefly describe the methods of microarray-based gene-expression profiling and highlight the technique's limitations, demonstrate the various levels at which both inter-experimental and intra-experimental variability can occur, summarize the suggestions that have been proposed to limit these problems, and describe the current status of genetic profiling in oncological practice.² The different steps involved in gene-expression-profiling using microarray technology and the major limitations of this method are depicted in Figure 1.

Figure 1 Key considerations when performing gene-expression profiling using microarray analysis.

The different steps in this process and their limitations are summarized. There are seven major steps in the development of a gene-expression signature using microarrays before it is ready for implementation as a test in clinical practice. The first step consists of defining the scientific question that should be answered by the output data. The accompanying hypothesis will help to construct a solid study design and to determine the type of samples and the applicable inclusion criteria (steps 2 and 3). Step 4 includes the actual microarray experiment by which the gene-expression data are obtained. The statistical analysis of the data generated in the experiment and the construction of the actual gene-expression profile is addressed in step 5. Step 6 includes the independent validation of the results, which is indispensable for the transition to step 7, the clinical implementation of the gene-expression profile. Figure modified with permission from reference 21 © (2002) Nature Publishing Group.²¹

Full figure and legend (71K)Figures & Tables indexDownload Power Point slide (278K)

DNA microarray technology

A DNA microarray is an ordered arrangement of equidistant microscopic DNA spots attached to a solid surface, such as a glass, plastic or silicon chip. Hybridization is performed using corresponding probes that recognize and attach to the solid support; these can be complementary DNAs (cDNAs), oligonucleotides of varying length, or genomic sequences that are either radioactively or fluorescently labeled (Figure 2). An array containing thousands of spots immobilized at predetermined locations can be generated by applying the DNA (e.g. cDNA or oligonucleotides) to the array using pins³ or inkjet technology,⁴ or by in situ photolithographic synthesis of oligonucleotides.⁵ For example, if inkjet technology is being used, an adapted color inkjet printer is used to synthesize oligonucleotides: each of the four color reservoirs is filled with a solution containing one of the four nucleotides from which DNA is made. The oligonucleotide sequences for each gene can then be assembled using a simple text file. Thus, the probes of a microarray are made of DNA and these probes are then used to detect the abundance of specific mRNA (transcripts) from the genes that correspond to the sequence of the probes on the array.

Figure 2 Gene-expression profiling using microarray analysis.

The main steps involved and limitations of microarray analysis will be encountered. Two stages can be distinguished in microarray experiments: a pre-microarray experiment phase (tissue handling) and the microarray experiment phase (labeling of the RNA). In a microarray experiment the glass slide with the labeled DNA of interest plays a key role. As can be seen, a solid surface (in this example a glass microscope slide) contains thousands of spots. Each spot contains a large number of identical DNA fragments. Fluorescently labeled RNA from the samples are subsequently hybridized to the arrays. In this way, the amount of DNA fragments per spot indicates the expression level of a gene. The expression level of thousands of fluorescently labeled genes or spots on one microscope slide can be visualized with a fluorescent scanner. For each gene on the array, the amount of fluorescently labeled RNA bound represents the expression level of that gene in the tumor sample. The intensity of fluorescent signal can be measured and used in the statistical analysis. For each spot, the DNA fragments are derived from one specific gene. Figure courtesy of Dr R Kerkhoven. Abbreviations: DHFR, dihydrofolate reductase; E2F1, E2F transcription factor 1; RB, retinoblastoma; SRC, sarcoma.

Full figure and legend (79K)Figures & Tables indexDownload Power Point slide (285K)

The concept of DNA or oligonucleotide arrays began in the mid-1980s. The first DNA arrays comprised nylon filters on a glass slide containing cDNA probes; these filters contained a much lower number of probes than the arrays that are currently used and were typically used with radioactively labeled targets. The introduction of pin-based robotic systems made it possible to dispense smaller volumes of DNA (approximately 150 microns) onto a glass slide, thus enabling a higher throughput system that represented one of the first microarrays.^{6, 7} A summary of the developments in microarray-based gene-expression profiling is provided in Table 2.

Table 2 The evolution of microarrays for gene-expression analysis over the years.⁶

Full tableFigures & Tables indexDownload Power Point slide (241K)

Microarray analysis of cDNAs spotted onto glass slides was developed at Stanford University and has been used to study the expression levels of large numbers of mRNAs in cell lines and tumors.⁸ In brief, mRNA is isolated from cells and reverse transcribed in the presence of red-fluorescent-labeled nucleotides. The resulting fluorescent cDNA is then mixed with a green-fluorescent-labeled reference cDNA and the mixture is hybridized to the microarray. The reference mRNA is usually prepared from a mixture of cell lines, tumor samples, or normal tissues. Using a fluorescent scanner, the fluorescence level is digitized, and for each cDNA on the microarray the level of gene expression, relative to the reference, is determined and transferred to a linked computer database.

Similar methodology is used for oligonucleotide-based arrays. With this system of microarray analysis, the expression of each gene in a sample is also measured relative to the expression of the same gene in the reference mRNA. The photolithographic synthesis of oligonucleotides is a procedure developed by Affymetrix (Santa Clara, CA); for each gene, several different oligonucleotides are present on the array.⁵ In this technique, the hybridization on the array platform is carried out using RNA from the sample to be analyzed, without the use of a reference RNA; instead, oligonucleotides containing one mismatch in their sequence are used to correct for background hybridization. An array containing 25,000 probes can provide information on the expression of all genes present in the genome. For many genes, multiple splice variants have been identified, and for sufficient array detection an increasing numbers of probes are being developed.

Gene-expression profiles can provide an enormous amount of information on cell function; however, it is important to realize that these profiles can only be used to interpret cellular changes that affect mRNA synthesis. After translation of mRNA into protein, many secondary protein modifications also play an important role in cell regulatory processes. Therefore, high-throughput analysis of proteins ('proteomics') will also contribute greatly to cancer research, an issue that was recently reviewed by Gulmann et al.⁹ Proteomic experiments can be tedious and have a lower throughput than RNA-based experiments;¹⁰ however, recent advances in proteomics have greatly increased the throughput of proteomic experiments.

Recently, small RNA molecules that do not code for proteins, such as microRNA, RNA interference, small interfering RNA (siRNA), and small modulatory RNA, have been discovered; these small RNAs have an important role in gene regulation.¹¹ Genes that regulate the expression of molecules such as microRNA are found in the non-coding regions (or introns) of the genome. Although these RNAs do not code for proteins, they can control gene expression at the post-transcriptional level by degrading or repressing mRNA,¹¹ thus influencing critical functions and biological processes.¹² Although microRNA consists of only 1–5% of human genes (which equates to about 1,000 microRNA genes), each microRNA might regulate as many as 200 genes, which implies that over one-third of the genes that encode for proteins are regulated by microRNAs.^{11, 13} MicroRNA expression profiles can be used to classify human cancers¹² because they reflect the developmental lineage and differentiation state of the tumors. Compared with normal tissue, microRNAs in tumors are generally downregulated and could thereby provide promising insight into tumor development and recurrence. For example, microRNA profiles are better predictors of metastatic origin for carcinomas of unknown primary (CUPs) than are conventional gene-expression signatures using mRNA.¹⁴

Statistical analysis of gene-expression data

Although not necessarily based on a gene-specific mechanistic hypothesis, good gene-expression profiling experiments should be planned and conducted with a clear objective, design and statistical analysis strategy.¹⁵ The main strategies used to identify categories of tumors by gene-expression profiling are unsupervised and supervised classification (Box 1). An important difference between these two methods is that for supervised classification clinical or pathologic information is used to find correlations with gene-expression patterns, whereas with unsupervised methods the tumors are grouped on the basis of their gene-expression pattern independent of the clinicopathological status.^{16, 17} A commonly used unsupervised method for examining gene-expression data is two-dimensional hierarchical cluster analysis. In this method, the tumors and genes are ordered according to similarities in their gene regulation, resulting in the clustering of tumors according to similarity in gene expression. This ordering often results in clustering of genes responsible for cellular processes, such as proliferation or inflammation. In this manner, several 'signatures' can be analyzed that reveal specific properties that the tumor cells and non-tumor cells contribute to the tumor mass. With time it is expected that an increasing number of signatures will be identified, making the correlation between gene-expression signatures in tumors and clinical behavior a powerful prognostic and predictive approach.

To find gene-expression patterns that can predict the clinical behavior of tumors (as in class comparison and class prediction), it is more appropriate to use a supervised classification that can distinguish specimens on the basis of predefined clinical and pathologic information, because the classes are already known. Several methods for supervised classification have been developed, and although these techniques were not specifically designed for the analysis of gene-expression data they have already shown their usefulness for this purpose.¹⁶ Fundamental to each of these techniques is the identification of the pattern of expression of a combination of genes that can predict tumor behavior (e.g. the risk for development of distant metastases, or responsiveness to specific treatments).

A third classification method is functional annotation of gene-expression algorithms based on data obtained from in vitro and in vivo experiments. An example of such an algorithm is the 'wound response signature'. Chang et al. compared the gene-expression profile of fibroblasts exposed to 10% serum with fibroblasts growing at low serum conditions; this gene-expression signature was termed 'wound response signature' because of the fact that fibroblasts are only exposed to serum in a wound.¹⁸ Based on the expression pattern in this 'wound response signature', breast carcinomas can also be divided into 'wound response signature activated' and 'wound response signature quiescent' tumors. Other gene-expression signatures discovered using in vitro experiments can in a similar way be applied to gene-expression data in malignant tumors. Chang et al.¹⁹ have combined three algorithms—the molecular subtypes or portraits (derived from unsupervised hierarchical clustering of breast carcinomas),²⁰ 'wound response signature',¹⁸ and the '70-gene signature'^{21, 22}—to reveal links between wound healing and cancer progression in breast cancer. A gene-expression profile based on the wound response signature provides the basis for prospectively assigning a prognostic score to patients and can be scaled to suit different clinical purposes. The wound response signature improves risk stratification independently of known clinicopathological risk factors such as tumor size, histologic tumor grade and nodal status, and previously established prognostic signatures based on unsupervised hierarchical clustering (i.e. 'molecular subtypes'²⁰) or supervised predictors of metastasis (i.e. the '70-gene signature'^{21, 22}).

Even when distinct subgroups are defined using gene-expression profiles, considerable heterogeneity within the broadly defined groups can confound the ability of the classifier to accurately predict outcomes for individual patients. One strategy by which to circumvent this heterogeneity is to combine or integrate multiple gene-expression patterns rather than a single gene-expression pattern, which provides a more powerful and robust classifier. Such defined sets of genes have also been termed 'metagenes' (Box 1); such metagenes can be combined with clinicopathological data to obtain optimal prognostic and predictive classifiers.²³ Moreover, this framework provides a mechanism for combining multiple forms of data—both genomic and clinical—to characterize most individual patients' gene profile and achieve the goals of personalized treatment.

Limitations of microarray technology

The applicabilty of microarrays in clinical practice

Microarrays have been used to study several tumor types, most notably breast,^{21, 22, 41, 42} ovary,⁴³ colon,⁴⁴ gastric, leukemias,^{45, 46} malignant lymphoma,⁴⁷ prostate,^{48, 49} lung,^{50, 51} CUP,⁵² and malignant melanoma.⁵³ Analysis and accurate clinical interpretation from the currently available data is a challenging task, as there are numerous inter-experimental variations that can significantly influence the interpretation of results. Proper statistical analysis and independent validation in large series of patients is the only way to address this dilemma. To understand the current status and relevance of the gene-expression profiles that have been developed, we have chosen to highlight a few important examples. Since current clinical trials evaluating the use of microarrays in the management of patients with breast cancer are ongoing, we have chosen to separate this subject for purposes of detailed discussion in a separate section.

Breast cancer

Results form gene-expression profiling studies for breast cancer are gradually being implemented in the clinic. The fact that prognostic factors have been used for the last 20 years to guide adjuvant systemic treatment of breast cancer means that gene-expression profiling can become an important adjunct to the known prognostic factors. For breast cancer, three relevant gene-expression profiles associated with prognosis have been identified: a 70-gene classifier,^{21, 22} a 21-gene signature,⁴¹ and a 76-gene expression profile.⁴² Table 3 shows the relevant similarities and differences between the three signatures. New additional signatures are being developed all the time, such as the recently identified signature from Pawitan and colleagues.⁶²

Table 3 Comparison between the three classified expression profiles: the 70-gene classifier,^{21, 22} the 76-gene expression profile,⁴² and the 21-gene signature.⁴¹

Full tableFigures & Tables indexDownload Power Point slide (326K)

In 2002, a 70-gene prognostic signature was identified by the Netherlands Cancer Institute in Amsterdam.^{21, 22} The investigators initially studied the frozen samples of 78 tumors on an Agilent^® platform (Agilent Technologies, Inc., Palo Alto, CA; 25,000 probes).²¹ Using supervised classification, they found that in patients under the age of 53 years the expression of a set of 70 genes best correlated to distant metastases as a first event in node-negative cancer. This 70-gene expression signature was subsequently validated in a second partially independent validation series of 295 breast cancer patients from the same institute.²² For the node-negative patients in the validation series, 61 tumor samples overlapped with the training series. The 70-gene signature also proved to correlate with the outcome of 144 node-positive breast cancer patients (10-year overall survival rates for 'poor' and 'good' prognosis signature were 59.5% and 92.0%, respectively).²² This 70-gene signature outperformed the traditional clinicopathological features, with a hazard ratio of 4.6 (95% CI 2.3–9.3, P < 0.001) in a multivariable analysis. A 'good' prognosis signature was present in 39% of the patients, and was associated with a 94.7% probability of freedom from distant metastases in the first 10 years, and an overall survival of 97.4% independent of nodal status. A 'poor' signature was present in 61% of patients, and at 10 years 60.5% of these patients remained free of distant metastasis and the overall survival was 74.1%. A gene-expression-profiling-based commercial test, called MammaPrint^®, is offered by Agendia (Amsterdam, The Netherlands).

In 2004, the American National Surgical Adjuvant Breast and Bowel Project (NSABP) in cooperation with the company Genomic Health identified a recurrence score including 21 genes that quantified the likelihood of distant recurrence in tamoxifen-treated patients with node-negative, estrogen-positive breast cancer.⁴¹ Fixed, paraffin-embedded tumor tissue was assessed and gene-expression measured,⁶³ resulting in the development of the Oncotype DX^® assay (Genomic Health, Inc., Redwood City, CA). The list of 21 genes (16 cancer-related genes and five controls) in this assay and the recurrence-score algorithm were designed by analyzing data from three independent preliminary studies involving 447 patients and 250 candidate genes identified in earlier (including microarray-based) studies. The 16 cancer-related genes were selected based primarily on the correlation of their performance with outcome in three trials—one of the trials is the NSABP B-20 trial, but the other two are not specified.⁴¹ Detailed descriptions of how the 16 genes were selected are lacking; the weight attributed to each of the 16 genes differs, but how the investigators arrived at these different weights has not been described in any detail.

To test the prognostic value of the recurrence score, RT-PCR was successfully used in 668 paraffin-embedded tumor blocks out of a larger study population of tamoxifen-treated patients in the NSABP B-14 study.⁶³ Tumor and patient characteristics in this subpopulation were similar to those of the total study population. The patients were of all ages and had tumors that were pT1–T2 estrogen receptor (ER)-positive, and were treated with tamoxifen. Using this 21-gene recurrence score, 338 patients (51%) had a low-risk, 22% an intermediate-risk, and 27% a high-risk profile. Multivariable analysis showed that the hazard ratio of this recurrence score was 2.81 (95% CI 1.70–4.64, P < 0.001). Clinicians in the US are using the 21-gene recurrence score in clinical practice based on validation studies described on the Genomic Health website (www.genomichealth.com). Only one study, however, has been published in the literature, and ironically this study failed to validate the recurrence-score algorithm.⁶⁴

In 2005, the Erasmus Medical Centre in Rotterdam in cooperation with the American company Veridex identified a signature of 76 genes that could identify node-negative breast cancer patients at high risk of distant recurrence who were therefore eligible for adjuvant systemic therapy.⁴² Using a retrospective study design, the Affymetrix chip U133A, which contains 22,000 genes, was used on frozen samples of 286 untreated node-negative T1–T3/4 breast cancer patients of all ages.⁴² This prognostic signature was identified using a training series of 171 tumors and consists of two separate profiles, one for ER-positive (60 genes) and one for ER-negative (16 genes) breast carcinomas. The gene-expression levels were analyzed using the log rank test, and validated in an independent set of 115 tumors without any overlap with the training set. After 60 months, the distant-metastasis-free survival in the 65% of patients with a 'poor' profile was 53%, and this number dropped to 49% after 80 months. A 'good' profile was observed in 35% of patients, the disease-free survival of whom was 93% after 60 months, and 88% after 80 months. After 60 months the patients with a 'poor' profile had an overall survival of 70%; the rate decreased to 63% after 80 months. Conversely, for patients with a 'good' profile the overall survival rates for 60 months and 80 months were 97% and 95%, respectively.

Although the 70-gene profile from the Amsterdam group and the 76-gene profile of the Rotterdam group have only three genes in common, most genes are involved in the same regulatory pathways. A reason why the overlap between these two signatures is small is that different microarray platforms were used (i.e. Agilent^® and Affymetrix^®). Among the differences between these two platforms is the fact that different DNA probes are used for the detection of the expression of specific mRNAs—the Agilent^® platform is a dual color hybridization using a reference RNA for each sample, whereas the Affimetrix^® platform only uses RNA from the tumor sample.

There are disadvantages in assessing gene-expression profiles identified from frozen tumor tissue. First, frozen tumor tissue is not widely available in clinical practice, and second, the assay is difficult to perform. These factors give the RT-PCR assay of Genomic Health an advantage over the other assays, since this assay can be performed more easily on readily available paraffin-embedded material. For samples that can be analyzed using paraffin-embedded tumor tissue, validation of the prognostic value can be established retrospectively from clinical trial material. For tests requiring frozen material, prospective studies are needed for validation. At present, the 70-gene signature^{21, 22} is being used in The Netherlands in the RASTER trial (Netherlands Cancer Institute collaboration with the Dutch Health Care Insurance Board), and soon the MINDACT trial (TRANSBIG consortium in collaboration with the European Organisation for Research and Treatment of Cancer) is expected to start using this signature.

All three signatures outperform classical clinicopathological parameters in a multivariable analysis, but insufficient independent validation has been performed to assess the reproducibility of the results.^{21, 22, 41, 42} It will be interesting to determine whether or not the three profiles categorize patients into the same prognostic groups. Although gene-expression profiling using microarray technology is very promising, there are still issues regarding its clinical applicability, and it is not yet known whether these profiles will genuinely improve the selection of those patients best eligible for adjuvant systemic treatment.

A key concern is when these gene-expression-based prognostic classifiers will be ready for clinical use. In the past 15 years there have been hundreds, even thousands, of studies attempting to identify novel prognostic factors in breast cancer, often employing molecular techniques. The factors used to predict prognosis and guide adjuvant treatment in most clinical guidelines are nodal status, tumor size, histologic grade, patient age and hormone-receptor status; recently, HER2 status has also been included. The reason that none of the factors has been implemented in clinical practice is that validation of the prognostic value often failed, and few prospective studies were properly planned. It appears that the gene-expression-based classifiers are more appealing to clinicians and patients, leading to a pressure to implement them in the clinic as soon as possible. Prospective validation of each of these classifiers in sufficiently large representative patient cohorts, however, is required before these tests can be used in the clinic. Premature use of these tests could lead to inappropriate advice on adjuvant systemic treatment, resulting in both undertreatment and overtreatment of patients with breast cancer.

Future directions

Microarray analysis is a rapidly evolving technology that has yielded enormous quantities of data, which are not yet fully understood. Leading scientific journals require investigators carrying out DNA microarray research to deposit their data in an appropriate international database,⁶⁵ following a set of guidelines (MIAME guidelines, or Minimum Information About a Microarray Experiment;² Table 4). This approach offers an opportunity to propose alternative analyses for these data. Michiels et al. took advantage of this repository to analyze different datasets from published studies of gene expression as predictors of cancer outcome.³⁴ They observed that most molecular signatures and misclassifications are unstable, and their main conclusion was that the list of genes included in a gene-expression profile (based on one training set and the proportion of misclassifications seen in one validation set) depends greatly on the selection of patients in the training sets.

Table 4 Useful websites with URLs for microarray-based research.

Full tableFigures & Tables indexDownload Power Point slide (260K)

With the recognition of the limitations of microarray analysis, but also of the measures that can be taken to reduce these limitations, and an understanding of the power of this technology, there is no reason why microarrays should not become a useful, if not essential, clinical tool. Until optimal gene-expression profiling is suitable for the clinic, the prognostic value of published microarray results in cancer studies should be interpreted with caution.

Technical issues should be addressed before this technology is translated into clinical practice. New methods of data analysis will increase the possibility of finding reliable prognostic and predictive profiles. Reproducibility and quality control are essential issues when using this technology as a standard assay in hospitals. In our opinion, the best way to improve analysis of microarray data is to start out with a clear study objective (e.g. class comparison, class prediction or class discovery study). If a prognostic or predictive profile is being built (class comparison or prediction), studies should be performed with statistical rigor and be reported clearly and with unbiased statistics.⁶⁶ Usage of the REMARK tumor guidelines for reporting of tumor marker studies could be a good start.⁶⁷ Currently, many studies have included only a small number of patients; however, for the future, adequate study size in combination with appropriate clinical design, adjustment for known predictors and proper validation will yield more reliable results and less overestimation of the predictive value of a classifier.^{38, 40} These features are essential in order for this highly promising technology to move from bench to bedside.

Simon et al. recommend that supervised methods rather than cluster analyses be used for class prediction and class comparison studies, as cluster analyses are less powerful than supervised methods for distinguishing predefined classes, and do not provide valid statistical identification of differentially expressed genes.¹⁵ If cross-validation is being used to estimate the prediction accuracy, the entire model-building process that includes selection of informative genes should be repeated for each cross-validation training set. Even after a successful cross-validation strategy is used, independent validation in an unselected patient cohort is indispensable.³⁴ If a separate dataset is being used for validation, it should be an unselected cohort that is sufficiently large to provide meaningful confidence intervals for prediction accuracy. Inadequate validation leads to the publication of overoptimistic results. Finally, the investigators should be urged not to make strong claims about the value of new prediction algorithms without comparing them against standard prediction methods.

Conclusion

Microarray technology is a very promising and quickly developing area of expertise that is leading to the discovery of numerous new signatures in a very short period of time. Although the use of gene-expression profiles in clinical practice is very appealing, we should be very cautious in interpreting all the high-throughput data and results before we start using any of the identified gene-expression signatures in daily clinical practice. Lack of frozen material from patients that are treated has been an important limiting factor preventing proper validation of reported gene-expression profiles. An ideal way to solve this problem is by incorporating gene-expression profiling into prospective randomized clinical trials. With proper validation and standardization of the microarray process, we expect that tests based on discoveries through gene-expression profiling will gradually enter the clinic in the next 10 years as an important adjunct to current clinical parameters.

Acknowledgments

We would like to thank L Wessel for advice and critical reading of the manuscript. We thank R Kerkhoven for his artwork for Figure 2.

References

1. Simon R et al. (2002) Design of studies using DNA microarrays. Genet Epidemiol 23: 21–36 | Article | PubMed | ISI |
2. Brazma A et al. (2001) Minimum information about a microarray experiment (MIAME)-toward standards for microarray data. Nat Genet 29: 365–371 | Article | PubMed | ISI | ChemPort |
3. Schena M et al. (1995) Quantitative monitoring of gene expression patterns with a complementary DNA microarray. Science 270: 467–470 | Article | PubMed | ISI | ChemPort |
4. Hughes TR et al. (2001) Expression profiling using microarrays fabricated by an ink-jet oligonucleotide synthesizer. Nat Biotechnol 19: 342–347 | Article | PubMed | ISI | ChemPort |
5. Lipshutz RJ et al. (1999) High density synthetic oligonucleotide arrays. Nat Genet 21: 20–24 | Article | PubMed | ISI | ChemPort |
6. Bertucci F et al. (1999) Sensitivity issues in DNA array-based expression measurements and performance of nylon microarrays for small samples. Hum Mol Genet 8: 1715–1722 | Article | PubMed | ISI | ChemPort |
7. Van de Goor TA (2005) A history of DNA microarrays. Pharm Discovery 5: 42–45 | ChemPort |
8. DeRisi J et al. (1996) Use of a cDNA microarray to analyse gene expression patterns in human cancer. Nat Genet 14: 457–460 | Article | PubMed | ISI | ChemPort |
9. Gulmann C et al. (2006) Array-based proteomics: mapping of protein circuitries for diagnostics, prognostics, and therapy guidance in cancer. J Pathol 208: 595–606 | Article | PubMed | ISI | ChemPort |
10. Lockhart DJ and Winzeler EA (2000) Genomics, gene expression and DNA arrays. Nature 405: 827–836 | Article | PubMed | ISI | ChemPort |
11. Chen CN et al. (2005) Gene expression profile predicts patient survival of gastric cancer after surgical resection. J Clin Oncol 23: 7286–7295 | Article | PubMed | ISI | ChemPort |
12. Lu J et al. (2005) MicroRNA expression profiles classify human cancers. Nature 435: 834–838 | Article | PubMed | ISI | ChemPort |
13. Meltzer PS (2005) Cancer genomics: small RNAs with big impacts. Nature 435: 745–746 | Article | PubMed | ChemPort |
14. Alizadeh AA et al. (2001) Towards a novel classification of human malignancies based on gene expression patterns. J Pathol 195: 41–52 | Article | PubMed | ISI | ChemPort |
15. Simon R et al. (2003) Pitfalls in the use of DNA microarray data for diagnostic and prognostic classification. J Natl Cancer Inst 95: 14–18 | Article | PubMed | ChemPort |
16. Knudsen S (2002) A Biologist's Guide to Analysis of DNA Microarray Data. New York: John Wiley & Sons
17. Eisen MB et al. (1998) Cluster analysis and display of genome-wide expression patterns. Proc Natl Acad Sci USA 95: 14863–14868 | Article | PubMed | ChemPort |
18. Chang HY et al. (2004) Gene expression signature of fibroblast serum response predicts human cancer progression: similarities between tumors and wounds. PLoS Biol 2: E7 | Article | PubMed |
19. Chang HY et al. (2005) Robustness, scalability, and integration of a wound-response gene expression signature in predicting breast cancer survival. Proc Natl Acad Sci USA 102: 3738–3743 | Article | PubMed | ChemPort |
20. Perou CM et al. (2000) Molecular portraits of human breast tumours. Nature 406: 747–752 | Article | PubMed | ISI | ChemPort |
21. van 't Veer LJ et al. (2002) Gene expression profiling predicts clinical outcome of breast cancer. Nature 415: 530–536 | Article | PubMed | ISI | ChemPort |
22. van de Vijver MJ et al. (2002) A gene-expression signature as a predictor of survival in breast cancer. N Engl J Med 347: 1999–2009 | Article | PubMed | ISI | ChemPort |
23. Nevins JR et al. (2003) Towards integrated clinico-genomic models for personalized medicine: combining gene expression signatures and clinical factors in breast cancer outcomes prediction. Hum Mol Genet 12: R153–R157 | Article | PubMed | ISI | ChemPort |
24. Van 't Veer LJ and De Jong D (2002) The microarray way to tailored cancer treatment. Nat Med 8: 13–14 | Article | PubMed | ChemPort |
25. Miller LD et al. (2002) Optimal gene expression analysis by microarrays. Cancer Cell 2: 353–361 | Article | PubMed | ISI | ChemPort |
26. Tu IP et al. (2004) A method for detecting and correcting feature misidentification on expression microarrays. BMC Genomics 5: 64 | Article | PubMed | ChemPort |
27. Forster T et al. (2003) Experiments using microarray technology: limitations and standard operating procedures. J Endocrinol 178: 195–204 | Article | PubMed | ChemPort |
28. Perez EA et al. (2004) Improving patient care through molecular diagnostics. Semin Oncol 31: 14–20 | Article | PubMed | ChemPort |
29. Paik S et al. (2005) Technology insight: application of molecular techniques to formalin-fixed paraffin-embedded tissues from breast cancer. Nat Clin Pract Oncol 2: 246–254 | Article | PubMed | ChemPort |
30. Brown I et al. (2003) From peas to "chips" - the new millennium of molecular biology: a primer for the surgeon. World J Surg Oncol 1: 21 | Article | PubMed |
31. Ramaswamy S and Golub TR (2002) DNA microarrays in clinical oncology. J Clin Oncol 20: 1932–1941 | PubMed | ISI | ChemPort |
32. Russo G et al. (2003) Advantages and limitations of microarray technology in human cancer. Oncogene 22: 6497–6507 | Article | PubMed | ChemPort |
33. Pusztai L and Hess KR (2004) Clinical trial design for microarray predictive marker discovery and assessment. Ann Oncol 15: 1731–1737 | Article | PubMed | ISI | ChemPort |
34. Michiels S et al. (2005) Prediction of cancer outcome with microarrays: a multiple random validation strategy. Lancet 365: 488–492 | Article | PubMed | ISI | ChemPort |
35. Jornsten R et al. (2005) DNA microarray data imputation and significance analysis of differential expression. Bioinformatics 21: 4155–4161 | Article | PubMed | ChemPort |
36. Ouyang M et al. (2004) Gaussian mixture clustering and imputation of microarray data. Bioinformatics 20: 917–923 | Article | PubMed | ChemPort |
37. Oba S et al. (2003) A Bayesian missing value estimation method for gene expression profile data. Bioinformatics 19: 2088–2096 | Article | PubMed | ChemPort |
38. Ioannidis JP et al. (2003) Genetic associations in large versus small studies: an empirical assessment. Lancet 361: 567–571 | Article | PubMed | ISI |
39. Simon R (2004) When is a genomic classifier ready for prime time? Nat Clin Pract Oncol 1: 4–5 | Article | PubMed |
40. Ntzani EE and Ioannidis JP (2003) Predictive ability of DNA microarrays for cancer outcomes and correlates: an empirical assessment. Lancet 362: 1439–1444 | Article | PubMed | ISI | ChemPort |
41. Paik S et al. (2004) A multigene assay to predict recurrence of tamoxifen-treated, node-negative breast cancer. N Engl J Med 351: 2817–2826 | Article | PubMed | ISI | ChemPort |
42. Wang Y et al. (2005) Gene expression profiles to predict distant metastasis of lymph-node-negative primary breast cancer. Lancet 365: 671–679 | PubMed | ISI | ChemPort |
43. Schaner ME et al. (2003) Gene expression patterns in ovarian carcinomas. Mol Biol Cell 14: 4376–4386 | Article | PubMed | ISI | ChemPort |
44. Eschrich S et al. (2005) Molecular staging for survival prediction of colorectal cancer patients. J Clin Oncol 23: 3526–3535 | Article | PubMed | ChemPort |
45. Yeoh EJ et al. (2002) Classification, subtype discovery, and prediction of outcome in pediatric acute lymphoblastic leukemia by gene expression profiling. Cancer Cell 1: 133–143 | Article | PubMed | ISI | ChemPort |
46. Ross DT et al. (2000) Systematic variation in gene expression patterns in human cancer cell lines. Nat Genet 24: 227–235 | Article | PubMed | ISI | ChemPort |
47. Alizadeh AA et al. (2000) Distinct types of diffuse large B-cell lymphoma identified by gene expression profiling. Nature 403: 503–511 | Article | PubMed | ISI | ChemPort |
48. Lapointe J et al. (2004) Gene expression profiling identifies clinically relevant subtypes of prostate cancer. Proc Natl Acad Sci USA 101: 811–816 | Article | PubMed | ChemPort |
49. Yu YP et al. (2004) Gene expression alterations in prostate cancer predicting tumor aggression and preceding development of malignancy. J Clin Oncol 22: 2790–2799 | Article | PubMed | ISI | ChemPort |
50. Garber ME et al. (2001) Diversity of gene expression in adenocarcinoma of the lung. Proc Natl Acad Sci USA 98: 13784–13789 | Article | PubMed | ChemPort |
51. Beer DG et al. (2002) Gene expression profiles predict survival of patients with lung adenocarcinoma. Nat Med 8: 816–824 | Article | PubMed | ISI | ChemPort |
52. Buckhaults P et al. (2003) Identifying tumor origin using a gene expression-based classification map. Cancer Res 63: 4144–4149 | PubMed | ISI | ChemPort |
53. Bittner M et al. (2000) Molecular classification of cutaneous malignant melanoma by gene expression profiling. Nature 406: 536–540 | Article | PubMed | ISI | ChemPort |
54. Ebert BL and Golub TR (2004) Genomic approaches to hematologic malignancies. Blood 104: 923–932 | Article | PubMed | ChemPort |
55. Armstrong SA et al. (2003) Inhibition of FLT3 in MLL: validation of a therapeutic target identified by gene expression based classification. Cancer Cell 3: 173–183 | Article | PubMed | ISI | ChemPort |
56. Lossos IS et al. (2003) HGAL is a novel interleukin-4-inducible gene that strongly predicts survival in diffuse large B-cell lymphoma. Blood 101: 433–440 | Article | PubMed | ChemPort |
57. Lossos IS et al. (2001) Expression of a single gene, BCL-6, strongly predicts survival in patients with diffuse large B-cell lymphoma. Blood 98: 945–951 | Article | PubMed | ISI | ChemPort |
58. Lossos IS et al. (2004) Prediction of survival in diffuse large-B-cell lymphoma based on the expression of six genes. N Engl J Med 350: 1828–1837 | Article | PubMed | ISI | ChemPort |
59. Etzioni R et al. (1998) Asymptomatic incidence and duration of prostate cancer. Am J Epidemiol 148: 775–785 | PubMed | ChemPort |
60. Kan T et al. (2004) Prediction of lymph node metastasis with use of artificial neural networks based on gene expression profiles in esophageal squamous cell carcinoma. Ann Surg Oncol 11: 1070–1078 | Article | PubMed |
61. Tamoto E et al. (2004) Gene expression profile changes correlated with tumor progression and lymph node metastasis in esophageal cancer. Clin Cancer Res 10: 3629–3638 | Article | PubMed | ISI | ChemPort |
62. Pawitan Y et al. (2005) Gene expression profiling spares early breast cancer patients from adjuvant therapy: derived and validated in two population-based cohorts. Breast Cancer Res 7: R953–R964 | Article | PubMed | ISI | ChemPort |
63. Cronin M et al. (2004) Measurement of gene expression in archival paraffin-embedded tissues: development and performance of a 92-gene reverse transcriptase-polymerase chain reaction assay. Am J Pathol 164: 35–42 | PubMed | ChemPort |
64. Esteva FJ et al. (2005) Prognostic role of a multigene reverse transcriptase-PCR assay in patients with node-negative breast cancer not receiving adjuvant systemic therapy. Clin Cancer Res 11: 3315–3319 | Article | PubMed | ISI | ChemPort |
65. Ball CA et al. (2002) A guide to microarray experiments-an open letter to the scientific journals. Lancet 360: 1019 | Article | ISI |
66. Golub TR et al. (1999) Molecular classification of cancer: class discovery and class prediction by gene expression monitoring. Science 286: 531–537 | Article | PubMed | ISI | ChemPort |
67. McShane LM et al. (2005) REporting recommendations for tumour MARKer prognostic studies (REMARK). Nat Clin Pract Oncol 2: 416–422 | Article | PubMed | ISI | ChemPort |
68. National Center for Biotechnology Information [http://www.ncbi.nlm.nih.gov]

Contact the journal about this article or read the Article Responses associated with this article.

Subject areas under which this article appears: Genetics/Genomics