Data integration methods validate a novel gene signature predicting patients' outcome

Motivation Cancer patients' outcome is written, in part, in the tumor microenvironment described by gene expression profile. Neuroblastoma (NB) tumor hypoxia is related to tumor aggressiveness and can be measured by gene expression. Different studies often measure gene expression profiles utilizing different platforms. For this reason, lacking of a sufficient large dataset is one of the main recurrent problems in microarray and RNAseq gene expression analysis. The consistent combination of data coming from different platforms can give new biological insights and stronger statistical support to the analysis. To this purpose, two approaches exist: meta-analysis and analysis by data integration. In the present work, the two approaches are compared in validating the prognostic ability of a novel neuroblastoma gene signature (NB-hop) measuring tumor hypoxia, which was specifically designed for outcome prediction. Methods Gene expression and clinical data of 498 and 283 patients were accessible in GEO at the number GSE62564 and in R2: microarray analysis visualization platform (http:// r2.amc.nl) as separate multi-platform datasets, respectively. Data are relative to the gene expression profiles of primary tumors measured by RNAseq via Illumina Hiseq 2000 and microarray via Affymetrix Exon ST1 platforms. Datasets were analyzed separately and then merged in a single crossplatform cohort. Merging was performed utilizing COMBAT algorithm implemented on the 'InSilico' merging package in R. Dendogram, PCA, PVCA, GOV, SOV validation methods were utilized to inspect the merging results. Leave- one-out cross validation utilizing Logic Learning Machine (LLM) carried out classification. We performed survival analysis by Kaplan-Meier method. Significance was assessed by log rank test statistics. P values smaller than 0.01 were considered significant. NB-hop gene signature was defined from the re-analysis of a previously published signature measuring tumor hypoxia but tailored to patient's outcome prediction [1]. Results The two datasets of 498 and 283 patients were firstly analyzed separately. Afterwards, the same datasets were merged into a single cohort of 781 patients. Inspection of merging results validated the effectiveness of the data integration approach. LLM stratified patients into two groups on the basis of NB-hop expression. NB-hop was highly accurate (Accuracy>75%) in predicting NB patients' outcome in the three datasets. Kaplan-Meier analysis showed a consistent significant low probability of survival in patients with highly hypoxic tumors (Log rank P<0.00001). Selected groups of high-risk patients were further stratified by NB-hop in statistically significant groups (Log rank P<0.01). Furthermore, outcome prediction analysis identified a homogeneous group of poor outcome patients defined by high NB-hop expression on all the three datasets. This group could be a new prototype of Ultra-High Risk patients' category. In conclusion, meta-analysis and analysis by data integration methods consistently validated the prognostic ability of the novel NB-hop gene signature in predicting patients' outcome. NB-hop identifies poor outcome patients that could benefit of hypoxia-targeted therapies.