Results: Our data-driven simulations and our application to survival prediction with eight breast cancer microarray datasets, suggest that standard cross-validation produces inflated discrimination accuracy for all algorithms considered, when compared to cross-study validation.
Gene Set Enrichment Analysis (GSEA) was run on the normalized, unfiltered microarray dataset as suggested in the tools implementation (http://software.broadinstitute.org/gsea/index.jsp) version 2.2.1) [ 34, 35] using the c5.all.symbolsols.symbols
GSEA was run on the normalized, unfiltered microarray dataset as suggested in the tools implementation (http://www.broadinstitute.org/gsea/ version 2.06), and using the c5 - GO gene sets collection of the Molecular Signatures Database (MSigDB) (http://www.broadinstitute.org/gsea/msigdb/ ).org/gsea/msigdb/
The superior performance of BEST in these synthetic datasets suggested that our algorithm worked well in the context of highly heterogeneous microarray data and was robust against moderately distorted data and sporadic outliers.
As noted above, inter-platform correlation of low-variance genes demonstrated wide variability within and across datasets, suggesting variable levels of noise in the microarray measurements.
Consistent with our results, a recent network analysis of microarray datasets from diabetes patients suggests PPAR signaling is the key underlying pathway in the pathogenesis of Type II diabetes [ 18].
This transformation serves both to aid the prioritization and to facilitate better comparability between microarray datasets, as it has been suggested that rank-based transformations of microarray data alleviate some of the issues associated with comparing cross-platform, cross-laboratory data (Irizarry et al., 2005).
Using simulated and real microarray datasets, Taylor and Tibshirani illustrated the performance of TSM, and suggested that TSM could be used to assess overall significance levels in microarray and other genetic studies with a number of hypotheses tests [24].
Our study suggests that local-least-squares-based methods are good choices to handle missing values for most of the microarray datasets.
The microarray datasets analyzed in this article are available in Gene Expression Omnibus, accession number GSE80748.
