Prostate cancer, among the most prevalent cancer types globally, exhibits marked heterogeneity and varying disease progression and clinical outcomes. Improved molecular subtyping is needed for patient stratification. Since prostate cancer has relatively few somatic point mutations, whole-transcriptome data instead offers a rich and relevant source of molecular data. We analyzed bulk tissue transcriptomes from four cohorts to characterize primary prostate cancer's cell type composition. A deconvolution of cell types was performed based on gene expression profiles. Patients with available multi-sample regional data from different tumor foci were analyzed for intrapatient heterogeneity. Three cell type composition subtypes were defined: T cells enriched (TCE), epithelial cells enriched (EPCE), and tumor-associated stromal cells enriched (TASCE). A machine learning model was developed to classify these subtypes and validated in three independent cohorts. The subtyping demonstrated a high correlation with established clinicopathological parameters (e.g., Gleason score, p-value < 0.05), and the classifier showed a promising ability to predict biochemical recurrence. Moreover, our analysis revealed that interfocal heterogeneity in patients with multifocal cancer significantly surpassed intrafocal heterogeneity (p-value < 0.05). In conclusion, this study provides a novel prostate cancer subgrouping based on cell type composition, with the TASCE subtype significantly associated with high biochemical recurrence risk.
Keywords: Cell type composition; Machine learning; Molecular subtyping; Multifocality; Prognostic biomarker; Prostate cancer.
Copyright © 2026. Published by Elsevier Inc.