DNA arrays provide quantitative and simultaneous measurements of the mRNA expression levels of thousands of genes in a biological sample (for a review, see [1,2]). This technology relies on the hybridisation of a labelled target derived from sample RNA to large sets of DNA fragments (the probes representing genes) arrayed on a solid support. The target is produced by reverse transcription of RNA and simultaneous labelling; it contains many different fragments of complementary DNA (cDNA) in various amounts in solution, corresponding to the numbers of copies of the original messenger RNA (mRNA) species. During the hybridisation, the amount of labelled cDNA that hybridises with its probe is proportional to its abundance in the original RNA sample. After washes and image acquisition, the signal present on each probe is automatically detected and quantified; it is proportional to the expression level of the concerned gene. Intensities are normalised and converted into expression levels that then are analysed.

Since hybridisation analysis produces an inordinate amount of data representing thousands of expression measurements, the validation, analysis, interpretation, display and storage of this raw information are not trivial steps. The quality of the data depends on well-controlled hybridisation conditions, such as a good quality RNA, the use of negative and positive controls, an adequate signal-to-noise ratio, and the widest possible dynamic range [2]. The analysis and interpretation of all of the validated data require the use of sophisticated analysis programs to extract the maximum meaningful information. In the past years, methods supervised such as neural networks and unsupervised such as the various clustering techniques have been applied [3].

Two different sorts of DNA arrays have been developed [2]. The first type uses arrays of cDNA clones robotically spotted on a support in the form of bacterial colonies or PCR products. The oldest versions are known as ‘macroarrays’, high-density filters that are made on large Nylon membranes (typically 100 cm²) with the target generally labelled radioactively. Macroarrays are flexible and compatible with the equipment present in most academic laboratories and can effectively assay sets of hundreds or a few thousands genes. We developed this approach in our laboratory few years ago [2,4,5]. ‘Microarrays’ are the miniaturised version: they can contain tens of thousands of PCR products on surfaces of a few square centimetres. The solid support is generally glass and targets are labelled with fluorescence, allowing simultaneous dual hybridisation of the test sample and of a reference RNA sample [6]. Microarrays can also be made on Nylon membranes: in this case targets are labelled with radioactivity [7] or colorimetry [8]. Hybridisation images of microarrays are obtained using high-resolution scanners or a sophisticated flat bed scanner.

The second type of DNA array was initially developed by Affymetrix (Santa Clara, CA, USA) and is known as an ‘oligonucleotide chip’. Hundreds of thousands of oligonucleotides are directly synthesised in situ on glass slides (a few square centimetres) [9]. For mRNA expression measurements, each gene is represented by a set of oligonucleotides (20–25 bp-long) complemented by a series of mismatched control oligonucleotides used to evaluate background. More recent implementations use longer and more specific oligonucleotides (60–80 bp-long) allowing each gene to be represented by a single sequence [10]. The target is labelled with fluorescence and the detection scanner is similar to that used for glass microarrays. Fig. 1 shows hybridisation images provided by these platforms.

By providing a global view of mRNA expression, DNA arrays provide a molecular biology tool capable of taking on the complexity and combinatorial nature of BC genetics. Their use is directed towards two major objectives [11]. The first, fundamental, objective is to better understand mammary oncogenesis; the identification of genes that drive disease progression will allow the development of new more specific therapeutics. The second, clinical, objective is to try to define a more accurate BC classification system. The clinical heterogeneity of disease is partially addressed by the current histoclinical parameters; in a number of cases, the treatment response and the clinical outcome vary widely between apparently similar tumours. The use of the comprehensive gene expression profiles of tumours may identify – based on the similarity in their expression profiles – new distinct molecular subgroups of tumours within histoclinically similar groups. Some recent publications suggest that these objectives may be reached within a few years.

3.1 DNA arrays and mammary oncogenesis

Key genes implicated in BC can be revealed by DNA array experiments comparing the molecular profiles of different stages of the diseases' development. Using Nylon arrays representing ∼200 candidate genes (Fig. 2A), we identified several genes differentially expressed between normal breast tissue and 34 primary breast tumours [16]. The genes ERBB2 and MUC1 were known to be involved in disease, whereas others, such as GATA3, had not yet been insinuated. We compared the expression profiles of node-negative tumours to tumours with massive axillary extension (10 or more positive nodes) and found a positive correlation between the expression level of ERBB2 and the number of tumour-involved nodes. Similar studies comparing primary tumours and metastases have been reported [17].

3.2 DNA arrays and prognostic classification of breast cancer

Such prognostic classifications were recently reported in localised BC [20,29], in locally advanced BC [26] and in multiple forms of disease [28]. In all these studies, including ours, no prognostic classification as accurate as expression profiles-based classification could be obtained using classical prognostic factors. In every case, the classification required expression measurements of some tens of genes that probably included potential molecular targets for new and more specific therapies. To further explore the validity of results we compared the lists of discriminator genes identified. Despite several different methodological aspects, 26 genes were found in at least two lists [30]. Reassuringly, some have a known prognostic value (e.g., ESR1, ERBB2) but most are not yet associated with prognosis but have functions that make them prime candidates for novel therapeutic targets.

All these studies thus revealed the great and promising potential of DNA arrays for capturing the true diversity of BC and refining the prognostic classification. But a number of issues need to be addressed before DNA arrays translate into clinical benefits [30]. We discuss here only the issue concerning the amount of starting RNA required that depends on the sensitivity of the chosen technique.

For less sensitive techniques, a greater quantity of starting material is required. One way to obtain a usable amount is to amplify the mRNA before labelling using linear amplification methods [31], but that remains coupled with a risk of modifying the relative abundances of individual sequence species in the target. However techniques are being improved, even though there is some bias in the amplification [32]. Another solution is to improve the intrinsic performances of current DNA arrays. Glass microarrays and oligonucleotide chips require a few micrograms of mRNA, while Nylon macroarray uses some micrograms of total RNA (5 μg in our studies). We compared the sensitivity of these three implementations without taking into account mRNA amplification methods [7]. Relating the minimal relative abundance, generally reported by authors as a sensitivity limit, to the target concentration, and ultimately to the size of the starting material, we showed that the sensitivity for Nylon macroarrays was in fact similar to other approaches: 20 to 30 million mRNA molecules are necessary in the sample to be detectable. We suggested that this sensitivity was mostly due to the high capacity of Nylon membranes (allowing large amounts of probe and correspondingly large signals) added to the excellent intrinsic sensitivity of radioactive detection. We then showed that Nylon microarrays combined with radioactive labelling provide sensitive expression measurements using submicrogram amounts of total RNA, i.e. 100 times less than the other methods. Fig. 3 shows images of Nylon microarray with 2000 cDNA clones hybridised with radioactive ³³P-labelled targets. It was first hybridised with a vector target corresponding to a sequence common to all spotted PCR products (Fig. 3A): this allows the estimation of the amount of target DNA accessible to hybridisation. It was then stripped and hybridised with a complex target made from only 0.5 μg of total RNA extracted from a cancer cell line (Fig. 3B). These developments are important for the application of DNA arrays in clinical oncology, because they will allow profiling of small tumour samples including biopsies, minimal residual disease or homogeneous material provided by microdissection.

DNA arrays provide a revolutionary tool to discover and analyse the complex molecular basis of cancer. Since their first appearance in the mid 1990s, many developments have allowed them to be accessed by more and more researchers. Many studies have confirmed their great potential interest in the scientific, biomedical and pharmaceutical fields of oncology. The clinical benefit for patients remains to be demonstrated. Combined with other emerging large-scale molecular analysis methods, they will provide great insights into our understanding of cancer. This would boost the development of new specific anticancer therapeutics and improve the current tumour taxonomy allowing the prediction with accuracy of the outcome of the disease, its sensitivity to a given treatment, and the delivery of treatments targeted to each individual tumour.