As DSGEO contains a non-uniform representation of various aspects of biology (for example certain tumor types such as breast and lung cancer were disproportionately represented), we applied Principal Component Analysis (PCA) as a dimensionality reduction procedure to minimize bias toward any particular lineage or cellular state. In this reduced eigenspace of 386 components (which explained 90% of the variance), cluster analysis was performed to identify tight clusters of commonly co-regulated transcripts. We applied an iterative peel-off procedure to select the centroids (Tseng and Wong, 2005). Specifically, at each iterative step in the tight clustering process, the k-means algorithm with K ranging 20-100 was applied repeatedly on 100 independent random subsamples each comprising 75% of the original data. This procedure yielded a consensus matrix that contained the proportion of trials a pair of genes were in the same cluster. Thresholding the consensus matrix yielded sets of genes that co-clustered in more than 80% of the trials. The genes belonging to the stable clusters were noted, excluded from the data and the procedure was repeated to identify additional clusters. Because high-dimensional data
