BACKGROUND: The molecular basis of the clinical heterogeneity of prostate cancer (PCa) is not well understood. OBJECTIVE: The purpose of our study was to identify and characterize genes in a clinically relevant gene expression signature in a subgroup of primary PCa positive for transmembrane protease, serine 2 (TMPRSS2)-v-ets erythroblastosis virus E26 oncogene homolog (avian) (ERG). DESIGN, SETTING, AND PARTICIPANTS: We studied gene expression profiles by unsupervised hierarchical clustering in 48 primary PCas from patients with a long clinical follow-up. Results were correlated with clinical outcome and validated in an independent patient cohort. Selected genes from a defined classifier were tested in vitro for biologic properties. INTERVENTION: Initial treatment of primary tumors was radical prostatectomy. OUTCOME MEASUREMENTS AND STATISTICAL ANALYSIS: Associations between clinical and histopathologic variables were evaluated by the Pearson χ(2) test, Mann-Whitney U test, or Kruskal-Wallis test, where appropriate. The log-rank test or Breslow method was used for statistical analysis of Kaplan-Meier survival curves. RESULTS AND LIMITATIONS: Most tumors that overexpressed ERG clustered separately from other primary PCas. No differences in any clinical end points between ERG-positive and ERG-negative cancers were detected. Importantly, within the ERG-positive samples, two subgroups were identified, which differed significantly in prostate-specific antigen recurrence-free survival, and cancer-specific and overall survival. From our findings, we defined a gene expression classifier of 36 genes. In a second, completely independent tumor set, the classifier also distinguished ERG-positive subgroups with different clinical outcome. In both patient cohorts, the classifier was not predictive in ERG-negative tumors. Biologic processes regulated by genes in the classifier included cell adhesion and bone remodeling. Tumor growth factor-β signaling was indicated as the main differing signaling pathway between the two ERG subgroups. In vitro biologic assays of two selected genes from the classifier (inhibin, beta A [INHBA] and cadherin 11, type 2, OB-cadherin (osteoblast) [CDH11]) supported a functional role in PCa progression. Possible multifocality and limited number of PCa samples can be limitations of the study. CONCLUSIONS: The classifier identified can contribute to prediction of tumor progression in ERG-positive primary prostate tumors and might be instrumental in therapy decisions.