Expression analysis could be used for the detection of novel regulatory networks and characterisation of primary tumours, leading to an improvement of cancer diagnosis. For this purpose metaGen has designed a custom GeneChip containing known tumour-associated genes (for example PTEN) and 1,500 cDNAs which were shown to be differentially expressed in EST libraries of several tumours by bioinformatics methods. Microdissected tissue needs to be used for accurate transcript profiling of solid tumours, but only low amounts of mRNA can be isolated from this source. Therefore validated and robust amplification techniques must be established. We compared the transcript profiles obtained from various amounts of mRNA as starting material from the human breast cancer cell line MDA-MB-231. As a standard for comparison we used 300 ng of poly(A)+-RNA which was labelled according to the Affymetrix protocol. For the amplification of 10 ng, 1 ng, 100 pg and 10 pg we applied a descibed aRNA technique1 until 300 ng of aRNA was obtained. This RNA was then labelled according to a modified Affymetrix protocol. For hybridisation and detection of the RNA we used the standard Affymetrix protocols. The GeneChips were analysed by the Affymetrix GeneChip software 3.1 and the coefficients for the average differences of the five experiments were calculated. For each of the four RNAs two cycles of preamplification with cDNA synthesis and in vitro transcription were necessary. We observed a correlation coefficient of 0.97 between the hybridisations with 10 ng and 1 ng, whereas the correlation coefficient for the hybridisations with 10 ng and 100 pg was 0.81. The correlation between the hybridsations with 300 ng (no preamplification) and 10 ng or 1 ng was 0.79 and 0.77, respectively. The correlation was weaker for the two other experiments. The preamplification of mRNA enables us to use low amounts of tissue. Based on the assumption that each cell contains roughly 0.5 pg mRNA, only 2,000 cells are required per hybridisation experiment. For better reproducibility expression levels should only be compared with samples that have undergone the same number of preamplification cycles.