OBJECTIVE The outcome of prostate cancer is highly unpredictable. To assess the dynamics of systemic disease and to identify patients at high risk for early relapse we followed the fate of disseminated tumor cells in bone marrow for up to 10 years and genetically analyzed such cells isolated at various stages of disease. METHODS Nine hundred bone marrow aspirates from 384 patients were stained using the monoclonal antibody A45-B/B3 directed against cytokeratins 8, 18, and 19. Log-rank statistics and Cox regression analysis were applied to determine the prognostic impact of positive cells detected before surgery (244 patients) and postoperatively (214 patients). Samples from primary tumors (n = 55) and single disseminated tumor cells (n = 100) were analyzed by comparative genomic hybridization. RESULTS Detection of cytokeratin-positive cells before surgery was the strongest independent risk factor for metastasis within 48 months (P < .001; relative risk [RR], 5.5; 95% CI, 2.4 to 12.9). In contrast, cytokeratin-positive cells detected 6 months to 10 years after radical prostatectomy were consistently present in bone marrow with a prevalence of approximately 20% but had no influence on disease outcome. Characteristic genotypes of cytokeratin-positive cells were selected at manifestation of metastasis. CONCLUSIONS Cytokeratin-positive cells in the bone marrow of prostate cancer patients are only prognostically relevant when detected before surgery. Because we could not identify significant genetic differences between pre- and postoperatively isolated tumor cells before manifestation of metastasis, we postulate the existence of perioperative stimuli that activate disseminated tumor cells. Patients with cytokeratin-positive cells in bone marrow before surgery may therefore benefit from adjuvant therapies.