Fig. 1: Cluster adherence of DLBCL samples.

Cluster adherence was determined using NMF clustering with the 304 DLBCL samples from the DFCI cohort, recapitulating the results from the original study, including five subtypes (C1, C2, C3, C4 and C5) and identical subtype assignment for all samples (left panel). Consensus clustering by resampling [21] illustrates the unstable character of NMF clustering, core patients (solid color, middle panel) and non-core patients (dashed color, right panel). To make this distinction, a stability score was determined by examining co-occurring sample pairs in the same subtype through 1000 iterations of NMF clustering. The heatmaps show patients by column and genomic features by row. Genomic feature colors in the heatmap indicate mutations (green), copy number losses (blue), copy number gains (red) and translocations (purple). DLBCL samples are clustered by subtype. The subtype bars on top indicate core DLBCL samples (colored bars) and non-core DLBCL samples (gray and dashed bars). Lymphgen annotation of the DFCI samples were taken from Wright et al. [7]. Left panel: heatmap of all 304 DLBCL samples from the DFCI cohort. Middle panel: heatmap of the 70% core samples with a high stability score and robust molecular subtypes. Right panel: heatmap of the 30% non-core samples with inconsistent subtype assignment throughout the clustering iterations.