Identification of DLBCL consensus clusters. (A) The left panel shows consensus matrices (clusters) produced by hierarchical clustering (HC), self-organizing maps (SOMs), and probabilistic clustering (PC). Clusters were generated using the top 5% of genes with the highest reproducibility across duplicate samples and largest variation across tumors and an approach that selected the most stable numbers of clusters with each algorithm (consensus clustering).⁷⁹  The right panel shows comparisons of the cluster assignments of different algorithms (PC versus HC, HC versus SOM, and PC versus SOM, respectively). Clusters are denoted as C[1,], C[2,], and C[3,]. The 3 clustering algorithms demonstrated excellent agreement with more than 84% of DLBCLs assigned to the same clusters by any 2 algorithms. (B) Expression profiles of the 3 DLBCL comprehensive clusters. The top 50 genes associated with each cluster are shown. Each column is a sample and each row is a gene. Color scale at bottom indicates relative expression and SDs from the mean. Red indicates high-level expression; blue, low-level expression. Adapted from Monti et al⁷⁹  with permission.