One goal for the gene expression profiling of cancer tissues is to identify signature genes that robustly distinguish different types or grades of tumors. Such signature genes would ideally provide a molecular basis for classification and also yield insight into the molecular events underlying different cancer phenotypes. This study applies a recently developed algorithm to identify not only single classifier genes but also gene sets (combinations) for use as glioma classifiers. Classifier genes identified by this algorithm are shown to be strong features by conservatively and collectively considering the misclassification errors of the feature sets. Applying this approach to a test set of 25 patients, we have identified the best single genes and two- to three-gene combinations for distinguishing four types of glioma: (a) oligodendroglioma; (b) anaplastic oligodendroglioma; (c) anaplastic astrocytoma; and (d) glioblastoma multiforme. Some of the identified genes, such as insulin-like growth factor-binding protein 2, have been confirmed to be associated with one of the tumor types. Using combinations of genes, the classification error rate can be significantly lowered. In many instances, neither of the individual genes of a two-gene set performs well as an accurate classifier, but the combination of the two genes forms a robust classifier with a small error rate. Two-gene and three-gene combinations thus provide robust classifiers possessing the potential to translate expression microarray results into diagnostic histopathological assays for clinical utilization.