Abstract Background: Gene expression profiling has been used to define subclasses of cancers and in the assembly of potential prognostic signatures. Molecular categories are commonly defined by supervised or unsupervised cluster analysis of mRNA levels. However, despite efforts to profile prostate cancer using transcriptome data, the genetic alterations and biological processes that contribute to clinical progression have made disease subtype categorization difficult. In addition, most studies are limited in statistical power, and therefore do not accurately reflect the heterogeneity of the disease in the general population. The goal of this study was to generate molecular profiles of prostate cancer using an unprecedentedly large transcriptome dataset. Methods: We compiled a very large compendium of 50 human prostate cancer transcriptome datasets containing over 4,000 human prostate cancer specimens, which we named the “Prostate Cancer Transcriptome Atlas” (PCTA). From this resource, 38 datasets, containing over 2,000 specimens, were combined into a single, normalized and batch-corrected dataset using the median-centering method. Molecular categories were sought using 29 published gene expression signature profiles, including castration-resistance (CRPC), oncogene activation, AR activation, AR variants, EZH2, FOXA1, TMPRSS2–ERG fusion, stemness, PTEN inactivation, polycomb complex (PRC2) repression, cell proliferation, epithelial-mesenchymal transition, and proneural and neuroendocrine differentiation. Signature profiles were computed by the weighted Z-score method. Results: Unsupervised clustering identified subsets of tumors manifesting consistent molecular features across many specimens, including a category showing strong AR signatures coupled in concert with significantly repressed PRC2 signatures. These features were mostly seen in high Gleason score (>7) or metastatic prostate cancer. Another major group was an “inverse” signature, showing activated PRC2 and repressed AR signatures. This stratification was independently validated using two datasets: (1) from 545 formalin-fixed paraffin-embedded (FFPE) tissue samples from primary prostate cancer from the Mayo Clinic (PLoS One 2013; 8(6):e66855) and 281 prostate cancers from a watchful-waiting cohort recruited in Sweden (BMC Med Genomics 2010;3:8). This analysis also allowed us to observe other, distinct variant classes, including activation of AR-variants and EZH2, high enrichment of stemness, and proneural and neuroendocrine differentiation. Conclusion: These results show that analyzing gene signatures using an integrated collection of transcriptome profiles from multiple platforms, and thereby significantly increasing sample size, allows the assignment of provisional disease categories that are either difficult to observe or not detected even in large profiling studies. Citation Format: Sungyong You, Jayoung Kim, Michael R. Freeman. Prostate cancer classification using a transcriptome atlas. [abstract]. In: Proceedings of the AACR Special Conference on Computational and Systems Biology of Cancer; Feb 8-11 2015; San Francisco, CA. Philadelphia (PA): AACR; Cancer Res 2015;75(22 Suppl 2):Abstract nr B1-63.