Fig. 3

MuAt prediction performance in PCAWG cancer genomes. a MuAt predictions stratified by the number of somatic mutations and whether the tumour type was correctly (solid colours) or incorrectly (cross-hatched colours) predicted (top-1). b Prediction accuracy (Y-axis) of MuAt and DNN by mutational burden (X-axis). MuAt results (blue points) show accuracies in repeated independent predictions (n=100; diamonds indicate mean accuracies). c Confusion matrix of the best-performing MuAt model. d Comparison of MuAt and DNN [20] accuracy (Y-axis) on sparse data with respect to the number of mutations in downsampled tumours (X-axis). Top-1/3/5 accuracies are shown. e Accuracy of MuAt with attention (w/ Att) and without (w/o Att), and with respect to the embedding dimensionality