Penalized likelihood methods for estimation and testing with genomic data



 DESCRIPTION (provided by applicant): In recent years the scientific community has acquired vast amounts of genomic data fueled by the promise of discovering the genetic and regulatory foundations of disease and phenotypic variation. This promise has not yet been fully realized, in part due to the limitations of our current statistical and computational tools. The problem of studying phenotypic associations with sequence variants is now well studied, but tools utilizing non-sequence data types and integrating multiple data sources are less well established. Many non-sequence data types possess spatial correlation with respect to genetic position - we might expect these data to follow a smooth function of position along the chromosome. In this research proposal we will develop adaptive methods for simultaneously estimating these functions or genomic profiles and discovering regions in which they are associated with clinically or biologically relevant outcomes. These methods are born out of a single penalized regression framework called Joint Adaptive Differential Estimation (JADE) and will be implemented in efficient, scalable algorithms. The suite of JADE methods will include binary and quantitative trait analysis, adaptive spatially varying clustering, and significance testing. These broad, flexible capabilities offer many possibilities for relating spatially structured genomic data typesto biological or clinical outcomes, or to other binary or quantitative genomic information such as gene expression levels. In our applications we will focus on DNA methylation data in a variety of healthy tissue types available from the Encyclopedia of DNA Elements (ENCODE) consortium and DNase I data in collaboration with a lab exploring in vivo changes in gene regulation associated with environmental changes.