Fig. 4

sysSVM2 benchmark on TCGA gastro-intestinal cancers. a Median receiver operating characteristic (ROC) curves across 657 gastro-intestinal (GI) samples from TCGA. Curves compare the ranks of canonical drivers to the rest of genes or to false positives. The median areas under the curve (AUCs) are also indicated. b Distribution of GI canonical drivers across the GI cohort. Lists of canonical drivers for each GI cancer type were obtained from NCG6 [11] and mapped to samples of the corresponding cancer type where they were damaged. Numbers of samples are indicated above each bar. Samples with five or more GI drivers did not require additional driver predictions. c Comparison of performance between sysSVM2 and four other driver detection methods. The set of unique drivers predicted by each approach were compared in terms of recall of GI canonical drivers, other canonical drivers (non-GI and outside the sysSVM2 training set) and false positives and proportion of novel predictions not previously associated with a cancer driver role. The number of genes in each category is reported in brackets. The recall of GI canonical drivers could not be assessed for sysSVM2 because these were part of the training set. They were however considered as drivers by default, rather than predicted by the algorithm. NA, not applicable. d Proportions of 657 GI samples left with no predicted drivers (left) or fewer than 5 predictions. The one sample left with fewer than 5 predictions by sysSVM2 (TCGA-FP-8210, stomach cancer) had four damaged genes overall