Melanomas are notoriously difficult to classify because of a lack of discrete clinical and pathological stages. Here, we show that primary uveal melanomas surprisingly cluster into two distinct molecular classes based on gene expression profile. Genes that discriminate class 1 (low-grade) from class 2 (high-grade) include highly significant clusters of down-regulated genes on chromosome 3 and up-regulated genes on chromosome 8q, which is consistent with previous cytogenetic studies. A three-gene signature allows biopsy-size tumor samples to be assigned accurately to tumor classes using either array or PCR platforms. Most importantly, this molecular classification strongly predicts metastatic death and outperforms other clinical and pathological prognostic indicators. These studies offer new insights into melanoma pathogenesis, and they provide a practical foundation for effective clinical predictive testing.