A Bayesian model-based approach to finding cell-type level associations in heterogeneous methylation samples
Daniel W Kennedy, ARC Centre of Excellence for Mathematical and Statistical Frontiers (ACEMS), School of Mathematical Sciences, Science and Engineering Faculty, Queensland University of Technology (QUT)
Monday, 18 September 2017, 3pm to 4pm
BDI LG Seminar Room 0
Epigenome-wide association studies are often performed using heterogeneous methylation samples, especially when there is no prior information as to which cell-types are disease associated. While much work has been done on estimating cell-type fractions and removing cell-type heterogeneity variation, relatively little work has been done on identifying cell-type specific variation in heterogeneous samples. In this talk I present a Bayesian model-based approach for making cell-type specific inferences in heterogeneous settings, by utilising a logistic transform to properly constrain parameters, and incorporating a prior knowledge of cell-type lineage via prior covariance structure. The approach was applied to the determination of sex-specific cell-type effects in methylation, where cell-type information was present as an independent verification of the results. The approach showed significant improvement in performance over previously used methods, particularly for detecting association in several rare cell-types. I outline current and future work on this problem, which leverages the flexibility of the Bayesian modelling approach by incorporating local methylation correlation and multiple data-types.