Fig. 4

Interpretation of cross-validated results. A All gliomas in the TCGA training set that were misclassified at least once over 1000 cross-validation trials. Only 3.4% of IDH-wildtype gliomas and 2.3% of IDH-mutant astrocytomas are misclassified in over 50% of trials. Misclassified IDH-mutant astrocytomas are disproportionately WHO grade 4. B Misclassified samples show regionality on a UMAP SCNA landscape: most misclassified IDH-mutant astrocytomas are found in an area of predominantly correctly classified IDH-wildtype tumors, and most misclassified IDH-mutant tumors are found in the area dominated by correctly classified IDH-mutant astrocytomas. Point-size of misclassified samples indicates the frequency of misclassification. C SHAP results of correctly classified IDH-wildtype tumors indicate that chromosome arm 10q drove correctly classified IDH-wildtype predictions. Each patient is assigned a point per chromosome arm. The magnitude of each point’s vertical coordinate is an indication of how influential the average SCNA status of the chromosome arm was to the classifier’s prediction. Positive vertical values favor IDH-wildtype predictions; negative values indicate a preference for IDH-mutant astrocytoma predictions. The color of each point corresponds to the average chromosome arm SCNA value: blue indicates loss, and red indicates gain. The magnitude of the bars drawn for each chromosome arm is the average chromosome arm SHAP score for all patients plotted; the bar’s color indicates the average SCNA state across all patients plotted. D SHAP results of correctly classified IDH-mutant astrocytomas; intact 10q drives these predictions. E A visualization of average SHAP score and average chromosome arm SCNA status over all correct and incorrect model predictions of IDH-wildtype and IDH-mutant astrocytomas shows the top 12 chromosome arms’ influence on model predictions. The height of groups of connected points orders each chromosome arm’s average influence (SHAP score) on correct predictions. The x-axis indicates the average extend of SCNA loss or gain of each chromosome-arm for the cohort of patients each point represents. The size of each point represents the average influence the chromosome arm has. Solid lines connect correct or incorrect predictions; dotted lines connect groups of the same IDH-subtype. Crossing dotted lines identify chromosome arms whose SCNA profile mimics the opposite IDH subtype when mistaken by our model