Overview of approaches to predict driver mutations. (a) Recurrent mutations that are found in more samples than would be expected by chance are good candidates for driver mutations. To identify such recurrent mutations, a statistical test is performed (see Table 2), which usually collapses all of the non-synonymous mutations in a gene into a binary mutation matrix that indicates the mutation status of a gene in each sample. (b) Assessing combinations of mutations overcomes some limitations of single-gene tests of recurrence. Three approaches to identify combinations of driver mutations are: (1) to identify recurrent mutations in predefined groups (such as pathways and protein complexes from databases); (2) to identify recurrent mutations in large protein-protein interaction networks; (3) de novo identification of combinations, without relying on a priori definition of gene sets. These approaches sequentially decrease the amount of prior information in the gene sets that are tested, thus allowing the discovery of novel combinations of driver mutations. However, the decrease in prior knowledge comes at the expense of a steep increase in the number of hypotheses considered, posing computational and statistical challenges. Different methods to identify combinations of driver mutations lie on different positions of the spectrum that represents the trade-off between prior knowledge and number of hypotheses tested.