Background: Microarray-based gene expression analysis is widely used in cancer research to discover molecular signatures for cancer classification and prediction. In addition to numerous independent profiling projects, a number of investigators have analyzed multiple published data sets for purposes of cross-study validation. However, the diverse microarray platforms and technical approaches make direct comparisons across studies difficult, and without means to identify aberrant data patterns, less than optimal. To address this issue, we previously developed an integrative correlation approach to systematically address agreement of gene expression measurements across studies, providing a basis for cross-study validation analysis. Here we generalize this methodology to provide a metric for evaluating the overall efficacy of preprocessing and cross-referencing, and explore optimal combinations of filtering and cross-referencing strategies. We operate in the context of validating prognostic breast cancer gene expression signatures on data reported by three different groups, each using a different platform.

Results: To evaluate overall cross-platform reproducibility in the context of a specific prediction problem, we suggest integrative association, that is the cross-study correlation of gene-specific measure of association with the phenotype predicted. Specifically, in this paper we use the correlation among the Cox proportional hazard coefficients for association of gene expression to relapse free survival (RFS). Gene filtering by integrative correlation to select reproducible genes emerged as the key factor to increase the integrative association, while alternative methods of gene cross-referencing and gene filtering proved only to modestly improve the overall reproducibility. Patient selection was another major factor affecting the validation process. In particular, in one of the studies considered, gene expression association with RFS varied across subsets of patients that differ by their ascertainment criteria. One of the subsets proved to be highly consistent with other studies, while others showed significantly lower consistency. Third, as expected, use of cluster-specific mean expression profiles in the Cox model yielded more generalizable results than expression data from individual genes. Finally, by using our approach we were able to validate the association between the breast cancer molecular classes proposed by Sorlie et al. and RFS.

Conclusions: This paper provides a simple, practical and comprehensive technique for measuring consistency of molecular classification results across microarray platforms, without requiring subjective judgments about membership of samples in putative clusters. This methodology will be of value in consistently typing breast and other cancers across different studies and platforms in the future. Although the tumor subtypes considered here have been previously validated by their proponents, this is the first independent validation, and the first to include the Affymetrix platform.


Bioinformatics | Computational Biology