Open Access Research

Dissection of a metastatic gene expression signature into distinct components

Paul Roepman1, Erica de Koning2, Dik van Leenen1, Roel A de Weger2, J Alain Kummer2, Piet J Slootweg2 and Frank CP Holstege1*

Author affiliations

1 Department of Physiological Chemistry, University Medical Center Utrecht, Universiteitsweg, Utrecht, the Netherlands

2 Department of Pathology, University Medical Center Utrecht, Heidelberglaan, Utrecht, the Netherlands

For all author emails, please log on.

Citation and License

Genome Biology 2006, 7:R117  doi:10.1186/gb-2006-7-12-r117

Published: 11 December 2006

Abstract

Background

Metastasis, the process whereby cancer cells spread, is in part caused by an incompletely understood interplay between cancer cells and the surrounding stroma. Gene expression studies typically analyze samples containing tumor cells and stroma. Samples with less than 50% tumor cells are generally excluded, thereby reducing the number of patients that can benefit from clinically relevant signatures.

Results

For a head-neck squamous cell carcinoma (HNSCC) primary tumor expression signature that predicts the presence of lymph node metastasis, we first show that reduced proportions of tumor cells results in decreased predictive accuracy. To determine the influence of stroma on the predictive signature and to investigate the interaction between tumor cells and the surrounding microenvironment, we used laser capture microdissection to divide the metastatic signature into six distinct components based on tumor versus stroma expression and on association with the metastatic phenotype. A strikingly skewed distribution of metastasis associated genes is revealed.

Conclusion

Dissection of predictive signatures into different components has implications for design of expression signatures and for our understanding of the metastatic process. Compared to primary tumors that have not formed metastases, primary HNSCC tumors that have metastasized are characterized by predominant down-regulation of tumor cell specific genes and exclusive up-regulation of stromal cell specific genes. The skewed distribution agrees with poor signature performance on samples that contain less than 50% tumor cells. Methods for reducing tumor composition bias that lead to greater predictive accuracy and an increase in the types of samples that can be included are presented.