Introduction

Overview

Diverse types of cancer genomics data are being collected widely and rapidly with the aim to systemically examine the origin and dynamics of different diseases. An important premise is that by integrating different types of genomics data, such as DNA copy number and RNA expression data, we will gain more knowledge about the underlying biological process. For example, high versus low correlation between a copy number aberration (CNA) for a gene marker and its abnormal RNA expression would indicate different disease mechanisms and therefore require different treatment strategies.

We propose a Bayesian model-based framework for the integration of different types of genomics data. We employ a mixture model (Parmigiani et al., 2002) for the observed expression data that defines latent indicators representing the differential expression status of each gene. By operating on the latent indicators, we effectively alleviate the high noise level in the original observed expression data. We integrate diverse types of genomics data through a regression of the latent variables across different data types. The regression model is naturally in agreement with the biological knowledge and allows for the easy incorporation of other covariates.

By definition, integration models must be able to borrow information from multiple genomic platforms, measured on the same patients and genes. For illustration purposes, we consider two of the most widely discussed genomic platforms: array comparative genomic hybridization (arrayCGH or aCGH), which measures DNA copy numbers, and expression microarrays, which measure RNA expression.