Many of the pediatric solid tumors categorized as small blue round-cell tumors (neuroblastoma, rhabdomyosarcoma, lymphoma and Ewing sarcoma) are difficult to distinguish morphologically, and routine immunohistochemistry may be insufficient to characterize them. Several specialized techniques are used to diagnose these cancers, including cytogenetics, interphase fluorescence in situ hybridization, the polymerase chain reaction with reverse transcription, and immunohistochemistry. However, despite the availability of these tests, poorly differentiated cancers can still pose a diagnostic dilemma. Gene expression profiling permits the simultaneous analysis of multiple markers, no one of which can perfectly classify any cancer. Based on our hypothesis that the gene expression profile is specific for each type, we have used complementary DNA microarrays containing 6,570 genes to investigate the expression profiles of 64 pediatric cancers. These included both tumor biopsy tissues (13 Ewing sarcomas and 11 rhabdomyosarcomas) and cell lines (10 Ewing sarcomas, 10 rhabdomyosarcomas, 12 neuroblastomas and 8 lymphomas). Using supervised clustering (F-statistics and weighted gene analysis), multidimensional scaling, hierarchical clustering, principal component analysis and layered perceptron prediction algorithms, we have identified 131 genes that accurately distinguish between these cancers. We have confirmed the accuracy of these findings in classification of these cancers on 12 blind samples. We have also identified several biologically relevant genes that are uniquely expressed in specific cancer types but that have not been previously associated with these diseases. Our results support the potential of cDNA microarrays as an efficient tool for developing a molecular taxonomy of cancer.