Our gliomasphere transcriptomes were classified into the three (Classical, Mes, and PN) clinically relevant TCGA sub-classifications as previous described (Laks et al., 2016 (link)). Briefly, the 173 core TCGA glioblastoma samples used in TCGA subclassifications of GBMs (Verhaak et al., 2010 (link)) were used to build our classification model. The TCGA unified gene expression data (across three microarray platforms: Affymetrix HuEx array, Affymetrix U133A array and Agilent 244K) were combined with our gliomasphere data from Affymetrix U133 plus 2.0 array, and utilizing the limma R package, they were normalized together (Smyth et al., 2005 (link)). Following batch effect normalization using the ComBat R package (http://statistics.byu.edu/johnson/ComBat/ ).
(Johnson et al., 2007 (link)), we used the LDA based centroid classification algorism (ClaNC) used by (Verhaak et al., 2010 (link)) to develop a 3-class centroid-based classifier from 38 Classical, 56 Mes, and 53 PN TCGA samples (Dabney, 2006 (link)), where the 26 TCGA neural samples were excluded. Only 789 of the of the 840 TCGA classifier genes were used to assign classifications to our gliomaspheres, due to limitations in gene name overlap between TCGA and our platforms.
(Johnson et al., 2007 (link)), we used the LDA based centroid classification algorism (ClaNC) used by (Verhaak et al., 2010 (link)) to develop a 3-class centroid-based classifier from 38 Classical, 56 Mes, and 53 PN TCGA samples (Dabney, 2006 (link)), where the 26 TCGA neural samples were excluded. Only 789 of the of the 840 TCGA classifier genes were used to assign classifications to our gliomaspheres, due to limitations in gene name overlap between TCGA and our platforms.