BACKGROUND: Setting a molecular classification for serous ovarian carcinoma from transcriptomic data has been proved to be troublesome due to the high specimen heterogeneity. Consequently, it made difficult the identification of survival prognosis gene markers. In this study, we proposed a tumor classification based on gene expression profiling and we derived a 10-gene survival Cox predictive model. METHODS: We selected eleven clinically-annotated microarray-based gene expression studies conducted in the last decade to collect gene expression values from 1'409 late-stage serous ovarian carcinoma specimens obtained from cytoreductive surgery. The 2'000 genes with the highest expression variability were selected for the analysis. The expression data were batch-harmonized using the method combat and then split into a training set (n=715) and two test sets (n=226 and n=468). The non-negative matrix factorization (NMF) clustering algorithm was used to derive four stable molecular clusters. Subsequently, twenty genes (five per cluster) were selected by penalized generalized linear models (GLMs) to build a predictive subtype classifier. The test set samples were used to validate the classifier. We also selected 10 genes by penalized Cox regressions to build a survival predictive model which performance was evaluated by log-rank tests and C-index computation using the test set samples. All computations and data analyses were conducted in R language. RESULTS: The NMF algorithm yielded four stable subtypes which molecular gene expression features were very similar to the clusters described by TCGA and obtained from a distinct tumor dataset. Thus, we attributed the same names as the TCGA subtypes: Differentiated, Immunoreactive, Mesenchymal and Proliferative. Survival differed significantly between the subtypes (log-rank test p-value of 0.00012) with the best vital prognosis for the Immunoreactive and the worst for the Mesenchymal group. Moreover, we observed 84% of concordance between the subtypes obtained using the NMF algorithm on the 2'000 genes and those using the 20-gene classifier; that supports the possibility to reduce the number of subtype gene markers for further classification systems. The 10-gene multivariable Cox regression survival model allowed segregating significantly good and poor prognosis samples from the largest test set (log-rank test p-value of 0.0002, n=468) and displayed a C-index of 0.543. However, the results on the smallest test set were not significant. Using the survival Cox model, we also identified predictive genes such as CXCL9, SERPINA1 or PTX3 which associated molecular pathways could be involved on cancer progression processes. CONCLUSION: The existence of the serous ovarian carcinoma subtypes previously described by TCGA was confirmed using gene expression profiling on a large multiplatform dataset. Several survival predictive genes and their associated pathways were identified and are valuable candidates for targeted therapies.