Additional file 4 of Systematic interrogation of mutation groupings reveals divergent downstream expression programs within key cancer genes

Additional file 4: Figure S3. Clustering subgrouping model coefficients reveals structure of mutation heterogeneity. We applied hierarchical clustering to examine the average regression model gene coefficients across all forty cross-validation folds for each of our subgrouping tasks. Subgroupings with task AUCs of below 0.7 were omitted, as were genes that did not have an absolute model coefficient ranked in the top five for any of the remaining tasks. Distances between subgroupings were computed by taking the inverse of the Spearman correlation across all gene coefficients; these were then used to cluster subgroupings into five groups. To facilitate presentation, here we only show these clusterings for subgroupings which did not have another subgrouping in the same cluster with a higher AUC and a Jaccard index of at least 0.9 with respect to the subgroupings’ mutated samples. The subgroupings with the highest AUC in each cluster are bolded, as is the gene-wide task. An asterisk is placed next to the AUCs of subgroupings with cv-significantly better performance than that of the gene-wide task. We include here these heatmaps for GATA3, TP53, and PIK3CA in METABRIC-(LumA) as well as KRAS in TCGA-LUAD. The corresponding figures for the remaining cases can be found at our data portal under Figures/S3 - Gene Coefficient Heatmaps. The names of these figures have the format (expr-source)__(cohort)__(gene)_auto-heatmap_Ridge.svg.