High-Dimensional Mixed-Effects Models for Multi-Omics Data
Clustered data are common in biomedical research, including multi-omics studies. Accounting for clustered data structure can increase estimation accuracy and improve power in testing. In this talk, I will present two high-dimensional mixed-effects models motivated by multiple types of clustered omics data: genomics, transcriptomics, and proteomics.
1) The first part of my talk is motivated by clustered proteomics data. I will discuss a high-dimensional multivariate mixed-effects model to test the association between multiple proteins from a functional pathway and triple negative breast cancer. A graphical lasso penalty is imposed to facilitate computation and capture the sparse biological correlation structure among proteins. Through simulations and real data analysis, I will demonstrate that the proposed method improves power as compared to existing univariate models.
2) The second part of the talk is inspired by emerging single-cell RNA-sequencing (scRNA-seq) data. With scRNA-seq data as references, we can deconvolve tissue expression into cell-type-specific expression. I will introduce an empirical Bayes method to deconvolve rich multi-measure tissue expression data, such as multiple regions of the brain tissue per subject. The estimated subject-level cell-type-specific expression enables analyses that are infeasible based on existing data. For instance, it provides novel findings in cell-type-specific expression quantitative trait loci (eQTLs) by integrating the deconvolved expression with genotype, as well as in cell-type-specific co-expression networks.
Jiebiao Wang is with the Department of Statistics and Data Science, Carnegie Mellon University.