Discovering Combinations of Genomic Aberrations Associated With Cancer

Publisher Website
Google Scholar
Abstract
This article introduces a model-based statistical methodology for the analysis of copy-number variations in cancer genomes measured by comparative genomic hybridization. The methodology allows one to infer combinations of genomic aberrations associated with the cancer phenotype. The stochastic model conjoins two features of cancer biology to infuse some context into an otherwise unsupervised learning problem. It asserts random genomic instability in a potential progenitor cell, followed by selection into a tumor of the descending cell lineage if the lineage experiences certain ensembles of genomic aberration. Disease heterogeneity is reflected in the possibility of a network containing multiple ensembles. The network of ensembles is an identifiable parameter. By forming the sampling model conditionally on selection, statistical dependencies (both positive and negative) can be induced between aberrations, and the model entails heterogeneity in the marginal rate of occurrence of aberrations. A double-Pólya distribution is introduced as a prior over the network of ensembles, and Markov chain Monte Carlo is developed to enable posterior computation. As an example, the methodology is used to reanalyze genomic aberrations from 116 renal cell carcinomas. It produces posterior probabilities that any given aberration is relevant to oncogenesis, posterior probabilities that pairs of aberrations reside in a common ensemble, and a point estimate of the network of ensembles. The methodology provides a model-based clustering of all measured aberrations according to these estimated ensembles and a model-based clustering of tumors according to the probable ensembles of genomic aberration that they have experienced. Although it is formulated here to analyze aberrations in cancer genomes, the instability-selection-network model may provide an approach to modeling dependence in correlated binary data on various biological systems. Limitations and possible extensions of the methodology are discussed.