Fig. 6: Diagnostic classification of childhood tumours.

a,b, Classification scores obtained on the test set, broken down by hierarchy level (a) and by a subset of representative pediatric tumor classes (b). These include accuracy (dark blue), AUCPR (orange), precision (blue), recall (green) and hierarchical similarity H (dashed gray). All averaged scores were calculated as micro (m) averages. The total reference population of each class is also shown as shaded bars (blue). c, Classification results obtained with the KiCS validation dataset. In blue is the fraction of confirmed diagnoses in the absence of reference samples; in cyan are confirmed diagnoses; in orange are samples that led to an update in diagnosis; and in gray are inconclusive cases. The internal circle fractions indicate samples with normal tissue contamination (empty circles) or low quality (dotted circles). d, Majority class assignment for patients with samples taken at multiple timepoints. Each sample is shown as a dot, with size proportional to the class probability. The full circle represents the majority class at the first hierarchical level; bottom half circles in transparency show further subtypes. On the right, the name of the transcriptional family assigned to the first sample is shown in short form, except for those where normal contamination was dominant, in which case the next available sample is used. Samples with multiple separate primaries are not shown (Supplementary Fig. 11). e, Classification probabilities for neuroblastoma samples, grouped by their majority assignment. Larger bars represent the assignment to classes to the first level of the hierarchy; thinner bars represent the confidence scores of neuroblastoma subtypes. Samples for which MYCN amplification was clinically identified in a pre-therapy sample are marked with a red star. Pre-therapy samples are marked with a gray caret. The lineage score for each sample and their reference group median are shown at the bottom as dots and dashed line, respectively. f, Class assignment probabilities for osteosarcoma samples, grouped by their majority assignment. Larger bars represent the assignment to the osteosarcoma or alternative classes; thinner bars represent the confidence scores of osteosarcoma subtypes. Pre-therapy samples are marked with a gray caret.