Abstract
The particular focus in this thesis is to provide a model-based framework for accurately predicting DNA methylation from genetic information using racially sparse public reposi- tory data. Epigenetic alterations are of great interest in cancer research but public repository data is limited in the information it provides; however, genetic data is more plentiful. The phenotype of interest is cervical cancer (CESC) in TCGA. Being able to generate such predictions would nicely complement other work that has generated predictions of gene expression from normal samples.
In the first chapter, I provide a thorough biological background, as well as an in-depth introduction to mixed model prediction. The second chapter of this thesis describes an application of the Classified Mixed Model Prediction (CMMP) ([Jiang et al., 2018]) method that enables accurate race-specific prediction of DNA methylation (DNAm) from genetic data lacking racial diversity such as in the TCGA CESC data. The predictive performance of the model is enhanced by combining different types of cancer data to increase data het- erogeneity and induce borrowing of information. These findings have been published in Genomics ([Rao et al., 2020]).
It has been known that dynamic methylation markers exhibit correlative patterns across the genome ([Teschendorff et al., 2014]). One limitation of applying CMMP in the case of CESC DNAm prediction is that the joint correlation structure among the methyla- tion outcomes was ignored, which motivates the multivariate version of the CMMP idea (mvCMMP). Therefore, I present mvCMMP in Chapter 3, an efficient multivariate classi- fied mixed model prediction approach that is derived from CMMP incorporating a flexible estimation method based on ASReml ([Gilmour et al., 1995]) for mixed effect prediction on multivariate outcomes. We demonstrate that mvCMMP addresses prediction accuracy in correlated outcomes through simulation as well as the TCGA CESC data.