A comprehensive analysis of nine prognostic signatures reveals the high classification instability in breast cancer.

Abstract

Abstract Abstract #2026 To determine the prognosis of breast cancer patients, clinical and pathological factors are currently employed. Gene expression micro-arrays offer new opportunities to determine individual prognosis. Publications have raised concerns about micro-arrays studies who have the potential to preclude their use in clinical routine. To improve the understanding of gene-expression classifiers we addressed the following issues: 1) Is the performance similar between independent classifiers? 2) Is proliferation a common biological theme that represents various signatures? 3) Are there other enriched pathways among signatures with prognostic ability?
 Methods:
 On 6 public datasets we applied the 76-gene signature; the Molecular subtypes; the Chromosomal Instability Signature; the Wound Signature; the Invasiveness Gene Signature; the Molecular Prognosis Index; and the Genomic Grade Index. Survival, predictive accuracy and overlap analyses were performed. We created enlarged signatures by including all probes with significant correlation to at least one of the genes in the original signatures. We gathered a collection of gene sets from four databases (GO, KEGG, Reactome, MSDB). For each signature, we evaluated whether specific gene sets (modules) are overrepresented. We tested the prognosis ability of each of them.
 Results:
 The survival and predictive accuracy analyses gave similar results for each of the 9 signatures. They all added significant information to a multivariate model including standard pathological and clinical criteria. Nevertheless, we showed that none of these signatures were able to identify good and poor prognosis patients when applied to samples with intrinsically poor prognosis features (Positive Lymph Node, Negative Estrogen Receptor, High Grade). Conversely they identified good and poor prognosis patients when applied to samples with intrinsically good prognosis features (Negative LN, Positive ER Low Grade). The overlap analysis showed a low agreement between the signatures. 50% of the samples had almost one discordant classification result out of the 9 classifiers tested. The intersection of the signatures revealed a set of proliferation genes. The signatures were build on 10 different gene ontology modules with prognostic ability.
 Conclusion:
 This study underlines the need of large prospective validation studies of gene expression signatures. Further computational intelligence and system biology studies would be held to determine the best way to use these classifiers in clinical routine. Citation Information: Cancer Res 2009;69(2 Suppl):Abstract nr 2026.