Model-based identification of outliers in gene expression

Detecting genes whose expression is altered in tumor cells is a major challenge in cancer research. Microarray technologies, which can measure the expression of thousands of genes simultaneously, has been widely used for this purpose. Because of the variety of molecular mechanisms that can lead to cancer, genes of interest are expected to be differentially expressed in a small subset of the samples, which, statistically, behave like “outliers”.

We developed an expression-based molecular classification method (POE) to discover novel biological classes and identify genes associated with them. The key idea of this method is that it models the gene expression using the latent categories that a gene is turned “on” or “off” compared to the baseline genes, and therefore estimates the probabilities of being differentially expressed. With the probabilities calculated, POE provides a numeric measure of how the genes are expressed, therefore helps researchers to determine whether a gene is outlier or not.

We compare our method with the previous approaches using simulation studies and a public dataset for transcriptional response of patients after radiation therapy. POE is proved to have more competitive power of finding the outliers and better control of the error rate.