Molecular voting for glioma classification reflecting heterogeneity in the continuum of cancer progression

Gliomas, the most common brain tumors, are generally categorized into two lineages (astrocytic and oligodendrocytic) and further classified as low-grade (astrocytoma and oligodendroglioma), mid-grade (anaplastic astrocytoma and anaplastic oligodendroglioma), and high-grade (glioblastoma multiforme) based on morphological features. A strict classification scheme has limitations because a specific glioma can be at any stage of the continuum of cancer progression and may contain mixed features. Thus, a more comprehensive classification based on molecular signatures may reflect the biological nature of specific tumors more accurately. In this study, we used microarray technology to profile the gene expression of 49 human brain tumors and applied the k-nearest neighbor algorithm for classification. We first trained the classification gene set with 19 of the most typical glioma cases and selected a set of genes that provide the lowest cross-validation classification error with k=5. We then applied this gene set to the 30 remaining cases, including several that do not belong to gliomas such as atypical meningioma. The results showed that not only does the algorithm correctly classify most of the gliomas, but the detailed voting results also provide more subtle information regarding the molecular similarities to neighboring classes. For atypical meningioma, the voting was equally split among the four classes, indicating a difficulty in placement of meningioma into the four classes of gliomas. Thus, the actual voting results, which are typically used only to decide the winning class label in k-nearest neighbor algorithms, provide a useful method for gaining deeper insight into the stage of a tumor in the continuum of cancer development.