On the statistical analysis of allelic-loss data

This paper concerns the statistical analysis of certain binary data arising in molecular studies of cancer. In allelic-loss experiments, tumour cell genomes are analysed at informative molecular marker loci to identify deleted chromosomal regions. The resulting binary data are used to infer properties of putative suppressor genes, genes involved in normal cell cycling. Various factors can complicate this inference, including background loss of heterozygosity, spatial (that is, within chromosome) dependence of the binary responses, non-informativeness of markers, covariates such as protein levels or tumour histology, heterogeneity of cells within tumours, and measurement error. We focus on the first three factors, discussing methods for statistical inference that separate background loss from significant loss. We outline the extension to other inferences, such as comparison questions and the relationship to covariates. Using characteristic features of tumourigenesis, we present a framework for the stochastic modelling of allelic-loss data, and build models within this framework; in particular, we propose a simple model that has chromosome breaks at locations of a Poisson process, and preferential selection cells with inactivated suppressor genes. We illustrate these methods on allelic-loss data from induced rat mammary tumours and human bladder cancers.