Merging microarray data, robust feature selection, and predicting prognosis in prostate cancer

Motivation: Individual microarray studies searching for prognostic biomarkers often have few samples and low statistical power; however, publicly accessible data sets make it possible to combine data across studies. Method: We present a novel approach for combining microarray data across institutions and platforms. We introduce a new algorithm, robust greedy feature selection (RGFS), to select predictive genes. Results: We combined two prostate cancer microarray data sets, confirmed the appropriateness of the approach with the Kolmogorov-Smirnov goodness-of-fit test, and built several predictive models. The best logistic regression model with stepwise forward selection used 7 genes and had a misclassification rate of 31%. Models that combined LDA with different feature selection algorithms had misclassification rates between 19% and 33%, and the sets of genes in the models varied substantially during cross-validation. When we combined RGFS with LDA, the best model used two genes and had a misclassification rate of 15%. Availability: Affymetrix U95Av2 array data are available at http://www.broad.mit.edu/cgi-bin/cancer/datasets.cgi. The cDNA microarray data are available through the Stanford Microarray Database (http://cmgm.stanford.edu/pbrown/). GeneLink software is freely available at http://bioinformatics.mdanderson.org/GeneLink/. DNA-Chip Analyzer software is publicly available at http://biosun1.harvard.edu/complab/dchip/.