Defense by Statistics student Huy Le of thesis titled "Differentially methylated regions in longitudinal, multivariate, and heterogeneous DNA methylation data".

Abstract: DNA methylation (DNAm) exhibits spatial dependence, whereby neighboring CpGs act in a coordinated fashion to exert their effects. For this reason, biologists are interested in identifying differentially methylated regions (DMRs), which are regions of spatially collaborative CpGs whose DNAm levels depend on a trait of interest. However, existing methods to infer DMRs are only applicable to univariate data and cannot be used in longitudinal, multi-trait, multiple cell type, and other multivariate DNAm data. To fill this gap, we develop the first method to infer DMRs in multivariate DNAm data. Our method is built upon a novel spatiotemporal model for trait-dependent DNAm, which we use to develop a new Bayesian inference paradigm to compare DMRs across time, traits, or cell types. Our work also offers a new definition of a DMR that is more congruent with biological intuition, as it considers both the strength of the association between the trait and the DNAm levels at the DMR's constituent CpGs, as well as the degree of the spatial dependence of its CpGs. We show the utility of our model by applying it to simulated and real multivariate DNAm data. We extend these ideas to heterogeneous univariate settings by developing a spatial empirical Bayes framework that learns multiple spatial dependence regimes genome-wide. This approach combines expectation propagation with mixture estimation to discover and infer region-wise DMR under heterogeneous spatial and error structures.

Advisor: Dr. Chris McKennan