Skip to main content
Fig. 1 | Genome Medicine

Fig. 1

From: Pan-cancer detection of driver genes at the single-patient resolution

Fig. 1

sysSVM approach for driver prioritisation. a Overview of sysSVM. Molecular (somatic SNVs, indels and mutation burden) and systems-level features (Additional file 2: Table S1) of damaged canonical drivers in the analysed samples are used for training. The best models of support vector machines (SVMs) with four kernels are selected using cross-validation and trained on the whole set of damaged canonical drivers. Finally, a combined score is used to prioritise driver genes in individual patients. The SVM implementation was generalised for optimal performance on a simulated cancer-agnostic dataset through data normalisation, parameter tuning and feature selection. b Generation of a simulated reference cohort from TCGA data. Values of damaging mutation burden and ploidy were randomly assigned to samples. Damaged genes were then extracted from real samples with similar values of damaging mutation burden (± 10% for mutations) and ploidy (± 0.1 for CNVs). Dots represent individual TCGA (orange) or simulated (yellow) samples. Red lines indicate average numbers of genes with damaging mutations or CNVs in TCGA samples, for each given values of damaging mutation burden or ploidy. c Frequencies of canonical drivers in real and simulated samples. Oncogene gain-of-function, tumour suppressor loss-of-function and both types of TP53 alterations were considered. d Gene sets used for sysSVM optimisation. The training set included oncogenes (OGs) and tumour suppressor genes (TSGs), as well as TP53. All other damaged genes were used for prediction and assessment. These included other canonical drivers (without a proven OG or TSG role), candidate cancer genes from published cancer sequencing screens, known false positives of established driver detection methods and the remaining damaged genes. Bars indicate the number of unique damaged genes across the reference cohort of 1000 simulated samples

Back to article page