Figure 5.

Multiparameter SOM analysis of human brain tumors permits patient stratification into clusters that predict clinical outcome. A, Clustering of SOM mappings. (left) Individual patient samples can be mapped to a SOM as described in Fig. 4A, where the color represents the frequency at which individual cells are assigned to each SOM unit. (center) To quantitatively compare SOM mappings from different patients, the neighborhood frequency (Neigh. Freq., the sum of the frequencies for a unit and its surrounding neighbors, Supplementary Fig. S12A) is calculated and arranged into a neighborhood frequency vector (NFV). (right) The NFVs for each patient sample are clustered, and the output is represented by a heatmap where red and green indicate relative high and low neighborhood frequencies, respectively. B, Clustering NFVs calculated from SOM mappings reveals distinct signatures of PI3K/Akt/mTOR pathway activity. NFVs representing the 19 tumor specimens were subjected to unsupervised hierarchical clustering, and the results were depicted using a heatmap where each row corresponds to a SOM unit and each column represents a tumor specimen. Red and green indicate relative high and low neighborhood frequencies, respectively. Tumor types are shown at right in the enlarged dendrogram for each patient (PA = pilocytic astrocytoma, D = dysembryoplastic neuroepithelial tumor, OA = oligoastrocytoma, OD = oligodendrioglioma, GS = gliosarcoma). Because the dendrogram indicated three predominant clusters, the average map and weighted signature for each cluster of patient samples were calculated. To assess which of the expression / phosphorylation levels contributed most to the multiparameter, SOM-based clustering, we performed an ANOVA analysis of the patient sample mean expression levels and found that EGFR was the most differentially expressed signaling / phospho-protein between the three SOM-based clusters (Supplementary Figure S12C). C, SOM-based clustering of tumor specimens correlates with clinical outcome measures. Kaplan-Meier curves for overall survival of the entire patient cohort (black) and Clusters I (red), II (green) and III (blue). Cluster II was significantly associated with an increased hazard of patient death (p < 0.05), whereas Cluster I was significantly associated with a decreased hazard of patient death (p < 0.001) (Supplementary Fig. S13, N.S. = not significant.)