Speaker: James Zou (http://people.fas.harvard.edu/~jzou/)

Affiliations: Microsoft Research, The Broad Institute, and Harvard University

Title: EWASHER: EPIGENOME-WIDE ASSOCIATION STUDIES WITHOUT THE NEED FOR CELL-TYPE COMPOSITION

Abstract: Epigenome-wide association studies (EWAS) seek to identify loci whose epigenetic changes are correlated with phenotype. Such analyses complement GWAS and can potentially yield key insights into disease etiology. EWAS faces many of the same challenges as GWAS in identifying the needles of true signal in the haystack that is the genome. Importantly, however, EWAS face an added challenge in that the epigenome can vary dramatically across different cell types. When cell-type composition differs between cases and controls, this leads to spurious associations that bury true associations. While the current approach to tackling this problem is to estimate the cell-type composition in each sample using laboriously collected and assayed “reference profiles” (measurements from purified cell type), we here propose to bypass the need for these reference samples altogether. Instead, we propose a method, EWASher, that automatically corrects for cell type heterogeneity, without cell-type compositions knowledge. We validate our method on a bronze standard methylation data set for which we have cell-type composition information, and demonstrate that EWASher performs as well as the state-of-art method, which explicitly uses cell type composition. We further validate the approach using extensive simulations. Finally, we apply EWASher to breast and colon cancer methylation data from the Cancer Genome Atlas (TCGA), where no cell-type composition is available. We find disease-relevant associations not obtainable by standard analysis. Our work is a step toward placing EWAS on a solid statistical footing comparable to that of GWAS, while removing the need to collect expensive and laborious reference profiles. Although this paper focuses on DNA methylation, we expect that EWASher can be generalized to other epigenetic data types. An accompanying Python-based software package for EWASher is available from http://mscompbio.codeplex.com/.