Purpose: The reliability of microarray-based cancer prognosis is questioned by Michiels et al. They reanalyzed seven studies published in the prominent journals as successful stories of microarray-based cancer prognosis and concluded that the originally reported assessments are over optimistic. We set to investigate the reality of microarrays for predicting cancer prognosis by using the same data sets with commonly accepted data analysis approaches.
Experiment design: Michiels et al.'s analysis protocol used a correlation-based feature selection method, split sample validation, and a nearest-centroid rule classifier. We examined their results through systematically replacing their analysis approaches with other commonly used methods as a parameter study. In addition, we applied a widely accepted permutation test in conjunction with 5-fold cross-validation to verify Michiels et al.'s findings.
Results: The stability of signature genes is likely obtained when a fold change-based feature selection method is applied. When cross-validation procedures are used to replace Michiels et al.'s split sample validation, only one of the seven studies yielded uninformative classifiers. The permutation test reveals that the confidence interval based on the split sample used in the Michiels et al.'s review is not a rigorous and robust approach to assess the validity of a classifier.
Conclusions: We concluded that the use of DNA microarrays for cancer prognosis can be demonstrated. We also stressed that caution should be exercised when a general conclusion is withdrawn based on a single statistical practice without alternative validation, which can leave a false impression and pessimistic perspective for emerging biomarker methodologies to advance cancer research.