Title: Searching for Gene Sets with Mutually Exclusive Mutations

Abstract: Cancer cells evolve through random somatic mutations. "Beneficial" mutations which disrupt key pathways (e.g. cell cycle regulation) are subject to natural selection. Multiple mutations may lead to the same "beneficial" effect, in which case there is no selective advantage to having more than one of these mutations. Hence we are interested in finding sets of genes whose mutations are approximately mutually exclusive (anti-co-occurring) within the TCGA Pancancer dataset. In principle, finding the best set is NP Hard. Nevertheless, we will show how a new Mutation anti-co-OCcurrence Algorithm (MOCA) provides an effective greedy search and testing algorithm with guaranteed control of the familywise error rate or false discovery rate, by combining some under-appreciated ideas from frequentist hypothesis testing. These ideas include: (a) A novel exact conditional test for the tendency of multiple sets to have a large/small union/intersection, which generalises Fisher's exact test of 2x2 tables. (b) Randomised hypothesis tests for discrete distributions. (c) Stouffer's method for combining p-values. (d) Weighted multiple hypothesis testing. A new approach to setting a-priori weights which generates additional implicit hypothesis tests is suggested, and allows us to preserve almost all statistical power when testing pairs despite introducing a combinatorially large number of additional hypotheses.