A recurrent data analysis problem arising in cancer biology is to separate sporadic genomic abnormalities from those which are not sporadic and thus are likely to have some biological significance. When multiple genetic alterations are necessary for tumor development, selection forces acting on the tumor cell lineage induce stochastic dependence upon the abnormalities presented by an observed tumor cell. I will discuss a simple framework for generating probability models for data, and will demonstrate this `instability-selection' framework for several data types including loss of heterozygosity and copy number variations measured by comparative genomic hybridization. The central contribution is a joint distribution for correlated binary data based on a simple network representation of tumorigenesis. I develop a Markov chain Monte Carlo algorithm to fit this model.
Copyright. All Rights Reserved.