A mixture modeling framework for differential analysis of high-throughput data

The inventions of microarray and next generation sequencing technologies have revolutionized research in genomics; platforms have led to massive amount of data in gene expression, methylation, and protein-DNA interactions. A common theme among a number of biological problems using high-throughput technologies is differential analysis. Despite the common theme, different data types have their own unique features, creating a "moving target" scenario. As such, methods specifically designed for one data type may not lead to satisfactory results when applied to another data type. To meet this challenge so that not only currently existing data types but also data from future problems, platforms, or experiments can be analyzed, we propose a mixture modeling framework that is flexible enough to automatically adapt to any moving target. More specifically, the approach considers several classes of mixture models and essentially provides a model-based procedure whose model is adaptive to the particular data being analyzed. We demonstrate the utility of the methodology by applying it to three types of real data: gene expression, methylation, and ChIP-seq. We also carried out simulations to gauge the performance and showed that the approach can be more efficient than any individual model without inflating type I error.