Accurate Outcome Prediction in Neuroblastoma Across Independent Data Sets Using a Multigene Signature

Purpose: Reliable prognostic stratification remains a challenge for cancer patients, especially for diseases with variable clinical course such as neuroblastoma. Although numerous studies have shown that outcome might be predicted using gene expression signatures, independent cross-platform validation is often lacking.

Experimental Design: Using eight independent studies comprising 933 neuroblastoma patients, a prognostic gene expression classifier was developed, trained, tested, and validated. The classifier was established based on reanalysis of four published studies with updated clinical information, reannotation of the probe sequences, common risk definition for training cases, and a single method for gene selection (prediction analysis of microarray) and classification (correlation analysis).

Results: Based on 250 training samples from four published microarray data sets, a correlation signature was built using 42 robust prognostic genes. The resulting classifier was validated on 351 patients from four independent and unpublished data sets and on 129 remaining test samples from the published studies. Patients with divergent outcome in the total cohort, as well as in the different risk groups, were accurately classified (log-rank P < 0.001 for overall and progression-free survival in the four independent data sets). Moreover, the 42-gene classifier was shown to be an independent predictor for survival (odds ratio, >5).

Conclusion: The strength of this 42-gene classifier is its small number of genes and its cross-platform validity in which it outperforms other published prognostic signatures. The robustness and accuracy of the classifier enables prospective assessment of neuroblastoma patient outcome. Most importantly, this gene selection procedure might be an example for development and validation of robust gene expression signatures in other cancer entities.