Human tumour tissue transplanted to and passed through immunodeficient mice as xenografts make powerful model systems to study tumour biology, in particular to investigate the dynamics of treatment responses, e.g. to chemotherapeutic agents. Before embarking on large-scale gene expression analysis of chemotherapy response in human sarcoma xenografts, we investigated the reproducibility of expression patterns derived from such samples. We compared expression profiles from tumours from the same or different mice and of various sizes, as well as central and peripheral parts of the same tumours. Twenty-three microarray hybridisations were performed on cDNA arrays representing 13000 genes, using direct labelling of target cDNAs. An ANOVA-based linear mixed-effects model was constructed, and variances of experimental and biological factors contributing to variability were estimated. With our labelling procedure used, the effect of switching the dyes was pronounced compared to all other factors. We detected a small variation in gene expression between two tumours in the same mouse as well as between tumours from different mice. Furthermore, central or peripheral position in the tumour had only moderate influence on the variability of the expression profiles. The biological variability was comparable to experimental variability caused by labelling, confirming the importance of both biological and technical replicates. We further analysed the data by pair-wise Fisher's linear discriminant method and identified genes that were significantly differentially expressed between samples taken from peripheral or central parts of the tumours. Finally, we evaluated the result of pooling biological samples to estimate the recommended number of arrays and hybridisations for microarray experiments in this model.