Bayesian Functional Data Analysis Over Dependent Regions and Its Application for Identification of Differentially Methylated Regions

• Activity Number: 581 - Advanced Cross-Disciplinary Statistical Methods in Statistical Genomics
• Type: Topic Contributed
• Date/Time: Thursday, August 6, 2020 : 3:00 PM to 4:50 PM
• Sponsor: Section on Statistics in Genomics and Genetics
• Author(s): Suvo Chatterjee* and Shrabanti Chowdhury and Duchwan Ryu and Fasil Tekola Ayele
• Companies: National Institute of Child Health and Development (NICHD)/ National Institutes of Health and Icahn school of Medicine at Mount Sinai and Northern Illinois University and NICHD/NIH
• Keywords: Functional Data Analysis; Dynamically weighted particle filter; Differentially Methylated Regions

Abstract

Bayesian functional data analysis provides flexible statistical inferences under harsh circumstances such as large volumes of data, considerable measurement errors and missing observations. Considering a sequence of genomic regions and functional data analysis on each region, where neighboring regions can be dependent, demanding computation is indispensable and analysis is sometimes infeasible for large number of regions. We model the dependency of neighboring regions using transition models and utilize a computationally efficient sequential Monte Carlo method to cope with the computational difficulties. We examine the efficiency of the proposed method through simulation and compare it with existing methods like Bumphunter. Simulation results indicate that our method has lower misclassification rates and computational time compared to methods that do not account for dependency and use classical multivariate analysis. We applied our method to identify differentially methylated regions in two independent cohorts of lung adenocarcinoma and birthweight. Our method was able to identify several functionally important genetic loci associated with the traits.