Additional file 1 of Systematic interrogation of mutation groupings reveals divergent downstream expression programs within key cancer genes

Additional file 1: Figure S1. Clustering of cohort transcriptomes reveals profiles consistent with molecular subtypes. We applied unsupervised learning to the expression data used for each cohort considered in this study in order to remove unwanted variation associated with molecular subtypes from our alteration divergence analysis. In conjunction with information on known molecular subtypes present in these cohorts, we identified cases such as METABRIC and TCGA-LGG in which these subtypes clearly overlapped with distinguishable transcriptomic profiles. This contrasted with cohorts such as TCGA-STAD in which subtypes were present but could not be unambiguously linked with unique transcriptomic profiles, and those like TCGA-LUSC in which neither molecular subtypes nor expression clusters were present. The counts of cohort samples with each subtype are listed in the plot legends. In cohorts where molecular subtypes were found to have identifiable transcriptomic profiles we created sub-cohorts that only included samples from a particular subtype or set of subtypes. Unsupervised learning on these sub-cohorts’ transcriptomes revealed that they did not exhibit the large-scale clusters of samples observed in the original cohorts and were thus much more suitable as input for our mutation classification pipelines. We include here these UMAP clusterings for the entire METABRIC cohort and the METABRIC-(LumA) cohort and for the remaining cohorts at our data portal under Figures/S1 - Cohort UMAP Clustering. The names of these figures have the format (expr-source)__(cohort)__UMAP_comps-0_1.svg.